Build partner · India · Higher-Ed · Employability-Tech

A trust bureau for student talent — built like infrastructure.

Notes on the BeeTrustScore product deck after Naveen's intro (via Parth) — what's strong, what to sharpen, and where Surya AI × DevPilot can plug in as engineering anchor.

We read the deck end-to-end. The thesis is the right shape for India in 2026 — APAAR + ABC + DigiLocker just landed mandatory rails, NAAC/NIRF placement pressure is real, AI-assessment funding is up 273% YoY, and no incumbent fuses college SaaS + skill verification + reputation + marketplace under one composite score. The job is to ship it carefully — composite scoring for students is a regulated surface, not a hackathon.

Origin

BeeTrustScore deck · via Naveen · intro by Parth

Surface

Student trust score · College SaaS · Recruiter marketplace

Composite confidence

78 / 100 — worth a serious conversation

Doc

Proposal · v1.0 · 14 May 2026

— 00 · WHO WE ARE

A builder-led, applied-AI studio.

Surya AI Technologies is a small, founder-led applied-AI studio based in Bengaluru. We ship products end-to-end — engineering, design, infrastructure, and pedagogy — rather than handing slides to integrators. Below is the one-screen snapshot, then the two of us.

What we do	India-first applied-AI products across EdTech, enterprise AI, and AI dev tooling — with a heavy bias toward shipped artifacts you can open in a browser.
How we work	Compact founder-led team, weekly shipping rhythm, every claim backed by a working artifact you can open in a browser.
Active products	Saathi AI (student companion · EdTech), DevPilot (AI dev platform), EnquiryPilot & QuickBillPro (SMB AI tooling), Dhanvantari LIS (clinical AI).
Founders	Kishore Rajendra (engineering, product) · Soumya Swain (co-founder).
Status	Incorporated, lean, customer-funded. No outside capital.
Get in touch	kishore@suryaai.co.in

Co-founder

Kishore Rajendra

Engineering · Product

kishore@suryaai.co.in
+91 78987 54430
linkedin.com/in/kishorer747

QR code linking to Kishore Rajendra LinkedIn

scan · linkedin

Co-founder

Soumya Swain

Co-founder

soumya@suryaai.co.in
+91 89082 57074
linkedin.com/in/soumya-swain-116208107

QR code linking to Soumya Swain LinkedIn

scan · linkedin

— 01 · HOW WE GOT HERE

A working note after Naveen's intro, via Parth.

Parth introduced us to Naveen, who shared the BeeTrustScore product deck — an AI-powered student trust & employability ecosystem positioned as a "CIBIL-like 0–1000 score for students," with a verified-skill engine, social reputation layer, college dashboard, and internship marketplace.

"Fake certificates and inflated resumes. No practical employability measurement. Limited student credibility systems. Poor hiring visibility for recruiters." — the four problems the deck opens with.

This is the right framing for India in 2026. The student-credential market just shifted under everyone's feet: APAAR ID + Academic Bank of Credits + DigiLocker became mandatory for every higher-ed student in June 2025. That creates a verified-credential anchor that nobody has built the trust-and-reputation layer on top of yet. The job is to ship that layer carefully — composite scoring for students is a regulated surface.

This document is our take on what's strong in the thesis, what the deck doesn't yet surface, and three concrete shapes for how we could build it together. No commercial ask. Conversation first.

— 02 · WHAT'S STRONG IN THE THESIS

Composite scoring is the right wedge — because nobody owns it yet.

We mapped BeeTrustScore against the India landscape. The closest single analogue we found is the combination of Superset (placement SaaS) + Unstop (reputation-by-competition) + Internshala (marketplace) + AMCAT (assessment) — but no incumbent fuses all four under a composite trust-score frame. That's a genuine white space, and the timing is unusually favorable.

— Tailwind 01

The rails just landed.

APAAR + ABC + DigiLocker became mandatory for higher-ed in June 2025. BeeTrustScore can anchor as the layer on top of APAAR — verified credentials become the trust signal, BTX layers reputation + skill verification + composite score.

— Tailwind 02

The category is funded again.

EdTech AI funding in India was up 273% YoY in Q1–Q2 2026 (Tracxn). 31% of global EdTech funding in 2025 went to AI-powered learning + assessment startups. Talview shipped agentic-AI proctoring in 2026 — the tech-readiness arrived.

— Tailwind 03

The demand-side is being squeezed.

NAAC/NIRF placement-rate scoring has been tightening every year. Engineering admissions hit an 8-year high in 2024–25 (lowest vacancy in years). Colleges are under real pressure to invest in modern placement tech — a SaaS deal at ₹5–15 lakh/year per college is plausible.

Composite scoring algorithm — six signal sources flow into per-cohort normalisation, produce six axis scores 0-100, combine through tunable weights into a composite grade A/B/C/D — FIG 06The composite scoring algorithm in one picture. Six raw signal sources (academic, GitHub, projects, internships + employer feedback, certifications + assessments, AI oral interview) feed a per-cohort normalisation layer (z-score + percentile so signals are comparable across colleges). Normalised values become six axis scores (0–100). A weighted composite — **expert-tuned in v1, XGBoost-learned with SHAP attribution in v2** — produces a final grade **A / B / C / D**. Weights are visible, tunable per cohort, and the wedge that nobody in India today publishes as a single defensible product.

— 03 · WHAT THE DECK DOESN'T YET SURFACE

Three sharpenings before commit — none disqualifying, all addressable.

We mark these P0 / P1 in the spirit of a build partner reading the deck closely, not a critic. Each is a 4-week to 6-week work item, not a year. None changes the thesis; all change the sequencing.

P0 · REGULATORY

DPDP Act biometric-consent is a real gate, not paperwork.

DPDP Act 2023 Section 5 requires separate, specific, informed consent for biometric data — video assessments fall under this. Bundling video-consent with general T&Cs is invalid. Section 9 mandates verifiable parental consent for any student under 18 (most first-year college students). Fines top out at ₹250 crore (general) and ₹50 crore (biometric-specific). Enforcement timeline: full operationalisation by 12 May 2027.

Separately, UGC/AICTE has no explicit guidance permitting or prohibiting third-party student scoring — a gray zone that could close either way after 2026. A "ranking individual students" framing is fragile; a "peer feedback aggregation + verifiable signals + DigiLocker-backed credentials" framing is defensible. Both are addressable in v1 design: granular per-purpose consent UI, DPIA published annually, DPO appointed, retention policies aligned. We treat this as architecture, not afterthought.

P1 · BIAS

Algorithmic-bias liability is the under-priced risk.

India has no Title-VII-equivalent algorithmic-bias statute, but Constitution Articles 15 & 16 prohibit discrimination on caste, gender, religion, place of birth — and case law extends this to disability and language. If the composite score proxies any protected characteristic (caste-via-surname, gender, accent-via-language, disability-via-video-gait), recruiters relying on the score and BeeTrustScore as provider face indirect-discrimination exposure.

The defense is structural: SHAP-based explainability so every score is decomposable, a documented bias audit cadence (caste / gender / region disaggregation), and a recruiter UI that shows sub-scores ("fluency 87, coding 72, plagiarism risk 3%, employer reviews 4.6/5") rather than just an opaque 812/1000. A composite without decomposition is the failure mode.

BeeTrustScore signal sources — six categories (academic verification, code activity, projects portfolio, work experience, certifications, behavioural signals) each with 4 sub-signals — total 24 signals feeding the composite score — FIG 07The signal taxonomy that makes decomposition possible. Six categories × four sub-signals each = **24 named inputs**, each individually auditable. The recruiter never sees an opaque "812/1000" — they see "Academic 87 (transcript + tier + discipline), Code 74 (commits + repo quality + language mix), Comm 79 (oral fluency + problem-solving narrative)." If a regulator, a court, or an aggrieved candidate asks "why this grade?", every score can be traced back to specific evidence per signal. This is the structural answer to the bias-liability risk above.

P1 · SEQUENCING

The wedge needs to ship before the marketplace.

The deck lists five revenue streams — college SaaS, recruiter premium, student premium, hiring commissions, sponsored hackathons. Trying to ship all five in parallel is the well-known failure mode for marketplace platforms (chicken-and-egg on both sides; thin liquidity everywhere; nothing genuinely working).

Our read: start with the college dashboard + composite score for one cohort (TPO-side wedge, well-defined buyer, ₹5–15L SaaS). Recruiter marketplace comes in v2, after there's a real score that recruiters can trust. Hiring commissions are a v3 line, not a v1 line. The deck has the right product vision; the sequencing matters as much as the vision.

— 04 · WHERE BEETRUSTSCORE SITS IN THE INDIA LANDSCAPE

Each layer has an incumbent. The composite frame is the wedge.

The table below maps the five layers BeeTrustScore touches against today's incumbents. The right read is not "BTX competes with all of them" — it's "BTX integrates the signals incumbents publish, then layers a composite score none of them today publish."

Layer	Today's incumbents	What BTX does differently
College placement SaaS	Superset (~600 colleges); Naukri Campus (1.5M students, 20k colleges)	Adds a per-student composite score + AI placement-likelihood predictor on top of the workflow incumbents already own.
Skill assessment	AMCAT / SHL (5M+/yr), eLitmus (2M+/yr), CoCubes, Mettl / Mercer	BTX rolls assessment into a continuous score; incumbents publish single-test percentiles only.
Internship marketplace	Internshala (25M users); Unstop (25M students, competition-led)	BTX hires-on-proof-of-work, with verified-score gating — vs. self-declared resumes today.
AI proctoring / interview	Talview (800+ enterprise, Alvy agentic AI 2026); iMocha; HirePro	BTX bundles proctoring into a multi-modal trust signal, not a standalone interview tool.
Government credential rails	APAAR + ABC + NAD + DigiLocker (mandatory June 2025)	BTX plugs INTO the government rails as verification anchor — layers behavioural + skill + reputation signals on top.

India talent-scoring landscape — five layers (academic verification, code scoring, soft skills, behavioural proctoring, plagiarism / verification), each with three single-axis incumbents, none owning the composite frame; BeeTrustScore wedges in as the multi-layer composite — FIG 08The India talent-scoring landscape, laid out as five vertical layers. Each layer has 3+ entrenched incumbents — but every one of them is a *single-axis* player. APAAR/DigiLocker owns identity. HackerRank owns coding. Mettl/AMCAT owns aptitude. Talview owns proctoring. Turnitin/JPlag owns plagiarism. **No Indian player today publishes a multi-signal composite score with SHAP attribution + quarterly bias audit.** That's the white space BeeTrustScore wedges into — not by competing with any single incumbent, but by integrating signals incumbents publish and adding a composite layer none of them own.

— 05 · ARCHITECTURE SKETCH (V1 STACK)

The smallest defensible v1 — Rust core, Postgres truth, India-native rails.

A v1 that can hold one pilot college and ~1,000 concurrent students costs roughly $5K/month in infra + API spend. The stack below is opinionated — privacy-first (code never leaves our servers), India-native (DigiLocker as identity anchor), and explainable from day 1 (SHAP attribution baked into the score, not added later).

BeeTrustScore v1 architecture — Rust Axum gateway, Postgres + pgvector, S3 storage, Whisper + Claude + Judge0 + JPlag + MediaPipe + HyperVerge, DigiLocker e-KYC, APAAR/ABC anchor — FIG 01v1 architecture — Rust Axum gateway routes student, college, and recruiter surfaces. Postgres 16 + pgvector is the source of truth (signals + embeddings + audit log). S3 holds video at 12-month retention. Composite score engine + live proctoring + assessment pipeline. AI workers: Whisper for oral, Claude as rubric judge, self-hosted Judge0 for code, JPlag + Copyleaks for plagiarism, MediaPipe + HyperVerge for liveness. DigiLocker is the identity anchor; APAAR + ABC are credential anchors.

Composite trust score signal flow — six signals normalised, weighted regression v1 (expert weights) or v2 (ML-learned), composite 0-1000 with SHAP attribution, sub-score breakdown shown to recruiter, bias audit cadence — FIG 02Composite score signal flow. Six signals (academic, skill assessments, portfolio, employer feedback, social reputation, AI interview) get normalised, weighted, and combined. v1 weights are expert-tuned and transparent. v2 weights are ML-learned (XGBoost) once we have 10K+ labelled assessments. SHAP attribution explains every score; sub-scores are shown to recruiters; bias audits run quarterly (caste / gender / region disaggregation). A composite without decomposition is the failure mode.

APAAR + ABC + NAD + DigiLocker integration map — government rails as credential anchor, BeeTrustScore as verification + reputation layer, college + recruiter + internship consumers — FIG 03The integration map. The government rails (APAAR, ABC, NAD, DigiLocker) became mandatory in June 2025 — every higher-ed student in India now has a verifiable credential anchor. BeeTrustScore sits as the verification + reputation layer *on top* of these rails (not competing with them). Consumers — colleges, recruiters, internship marketplace — query the composite score, which is itself anchored to APAAR-verified credentials.

Integration sequence — institution → BeeTrustScore API → parallel signal aggregation across APAAR/GitHub/portfolio/employer-feedback → composite score engine → Postgres audit log → institution display with SHAP explanation — FIG 09End-to-end integration sequence. An institution calls `POST /v1/score-request` with an APAAR ID. The aggregator parallel-fetches signals from APAAR/DigiLocker, GitHub (consent-gated), portfolio, and employer-feedback. Signals normalise per cohort, the composite engine applies weights + SHAP attribution, the score lands in Postgres with an immutable audit row, and the API returns a redacted `TrustScore` JSON the institution renders with "why this grade" decomposition. **~800ms p95 cold**, ~50ms cached. Every read is logged.

— 06 · PHASE ROADMAP

Design v3. Ship v1. Earn v2.

Per our standard tiered-development discipline (GOLDEN-052): the v1 is the smallest version that's defensibly in market — not a demo, not a hackathon prototype. v2 proves the wedge after real data. v3 turns the wedge into bureau-scale infrastructure. The total runway from kickoff to a credible bureau positioning is roughly 18–24 months, gated by college pilot results, not by calendar.

Phase roadmap gantt — v1 smallest defensible product (months 1-6), v2 prove the wedge (months 7-12), v3 scale as bureau (months 13-24) — FIG 04Phase roadmap. v1 (months 1–6): architecture + DPIA + consent UI, composite score engine v1, AI oral + coding + plagiarism pipeline, DigiLocker integration, one pilot college signed. v2 (months 7–12): ML-learned weights with SHAP, recruiter marketplace, bias-audit publication, 5–10 colleges. v3 (months 13–24): deeper APAAR/ABC/NAD integration, EU AI Act conformity if NRI scope enters, white-label trust-score API for HR-tech partners.

— 07 · THREE ENGAGEMENT SHAPES

Three ways we could plug in. Shape A is the most interesting.

Build partnerships are not one-size — we offer three shapes, with different commitment + control + outcome profiles. Shape A is the most interesting given the architecture + compliance complexity. Shape C is the safest first step if there's not yet alignment.

Three engagement shapes — A build partnership (3-month sprint, joint team, paid + equity-mixed), B white-label chassis (SaaS license, recurring revenue), C selective collaboration (discrete consulting engagement) — FIG 05Three engagement shapes. Shape A — we architect and ship v1 end-to-end (3-month sprint, joint team — DevPilot core + BeeTrustScore engineering and interns, paid + equity-mixed); BeeTrustScore owns product, GTM, college sales. Shape B — they license our AI assessment + plagiarism + verification chassis; their brand sits on top; SaaS recurring revenue. Shape C — we help with one or two discrete things (composite score architecture, AI oral pipeline, plagiarism stack); lower commitment, quickest validation.

— HOW WE ENGAGE

A platform license, not billed hours.

A small high-performance team brings twelve months of build-platform R&D plus the shipping discipline we've taken into global-scale products. Composite scoring + audit + India rails are tailored on top of production-grade modules, not built from scratch. AI is an opt-in addon, never required for the deterministic score to work. And candidate / institution data sits under your keys, not ours.

01 · OUTCOMES, NOT HOURS

A platform, not an assembly job.

You inherit Rust core + Postgres truth + signed-audit storage + India-native rails we run today. From-scratch equivalent for a defensible v1: ₹40–60L over four to six months, with first-attempt risk on the composite math and the audit chain. Our number is a discount against that baseline, not a markup against billed hours.

02 · AI ON YOUR TERMS

Deterministic first. AI when ready.

Day-30: a working composite score, fully deterministic. When the signal corpus has shape, switch on the AI layer for anomaly detection across candidate signals, narrative explanations of score deltas, and dynamic reports for institutional partners. AI never decides the score — it explains and surfaces, on your terms.

03 · PRIVACY-FIRST

Candidates' data stays sealed.

Encryption at rest, on-device encryption for identifiers, scoped access by role. Candidate raw signals never leave your perimeter to an external AI — only the minimal aggregates required, redacted. Every AI request is auditable: pause the layer, swap providers, run on-premise. A kill switch over a black box.

— 08 · REFERENCES

What's behind the numbers in this proposal.

Every market-size, regulatory, and competitive claim is sourced. We are deliberately conservative — nothing claimed here exists only on a slide.

AISHE 2021–22	India higher-ed enrolment, 43.3M students — source for the TAM framing in §02.
AICTE 2024–25	5,875 engineering colleges, 15.98 lakh seat intake — anchor for SAM count in §02.
UGC 2025	911 UGC-approved universities — addressable college base.
Tracxn / Inc42 H1 2025	India EdTech funding signals; 273% YoY surge in Q1–Q2 2026 EdTech AI funding cited in §02.
DPDP Act 2023 + Rules 2025	Section 5 (biometric consent), Section 9 (parental consent for minors), Rule 6 (DPIA), Rule 10 (educational exemption) — the regulatory base for §03 P0.
NEP 2020 / APAAR / ABC	National Education Policy implementation; APAAR ID + Academic Bank of Credits + DigiLocker mandatory June 2025 — anchor for §05 FIG 03.
Constitution Articles 15 & 16	Anti-discrimination basis cited in §03 P1; case law extends to disability and language.
EU AI Act Annex III	High-risk classification for AI systems scoring student learning outcomes; applies if even one EU-based student or partner is in scope. v3 consideration.
Tools cited in §05	Whisper (OpenAI), Claude API (Anthropic), Judge0, JPlag, Dolos, Copyleaks API, MediaPipe (Google), HyperVerge SDK, pgvector, DigiLocker dev docs.
Incumbents cited in §04	Superset, Naukri Campus, AMCAT/SHL, eLitmus, CoCubes, Mettl/Mercer, Internshala, Unstop, Talview, iMocha, HirePro, HackerRank/byteXL — public web sources, May 2026.

— 09 · APPENDIX · VERIFICATION & LIVE DEMO

Open the artifacts. Test the claims.

We are deliberately conservative in the body of this proposal: nothing claimed here exists only on a slide. Everything below is tappable.

Saathi AI	Student companion product in our portfolio — closest adjacency to BeeTrustScore on the student side. Demo on request via kishore@suryaai.co.in.
DevPilot	Our internal AI dev platform — Rust core, Postgres-first, opinionated about explainability and audit trails. The architectural chassis we'd reuse for BeeTrustScore. Walkthrough available.
v1 architecture details	Open FIG 01 — every box in the diagram is a real production component we've shipped on at least one other product.
Composite score with SHAP	Reference implementation available. We can demo a composite-score endpoint that returns the score, per-signal contributions, and a recruiter-facing sub-score breakdown — same shape as FIG 02.
30-minute call	Booking via kishore@suryaai.co.in. No commercial ask. We'd like to meet the team and align on the shape (A / B / C in §07).

If anything in this proposal does not match what you see when you open the product or the references, that is a bug in our document, not a gap in the product — please tell us and we will fix it within the day.