Prakhar Shukla

Full-Stack Engineer · Agentic AI Systems · IEEE Published Researcher

I build AI systems you can actually trust - fraud defense, hallucination detection, and research integrity - with held-out metrics, honest baselines, and production hardening.

Most builders make models faster. I make their outputs trustworthy - verifiable, adversarially tested, and safe to put next to money. Five production systems, one thesis: models propose, deterministic code decides.

                     ┌──────────────────────────────────────────────┐
                     │             THE AI TRUST LAYER               │
                     │      models propose · code decides           │
                     └───────────────────────┬──────────────────────┘
          ┌──────────────────────┬───────────┴───────────┬──────────────────────┐
          ▼                      ▼                        ▼                      ▼
   ┌─────────────┐        ┌─────────────┐          ┌─────────────┐       ┌─────────────┐
   │  SENTINEL   │        │  GATEHOUSE  │          │ SIGNALGATE  │       │ TRUTHLAYER  │
   │ cross-mrcht │        │ household   │          │ quant       │       │ LLM output  │
   │ fraud graph │        │ scam agent  │          │ integrity   │       │ firewall    │
   ├─────────────┤        ├─────────────┤          ├─────────────┤       ├─────────────┤
   │ 0.857 F1    │        │ 1.00 / 1.00 │          │ 0.925 catch │       │ 95.33% prec │
   │ p50 2.3 ms  │        │ 0.0% f-gate │          │ p = 0.00029 │       │ 5-signal    │
   │ +₹38.6k/1k  │        │ $0.00026/cs │          │ byte-repro  │       │ McNemar✓    │
   └─────────────┘        └─────────────┘          └─────────────┘       └─────────────┘
   defense-only           real-household            recommends,           calibrated
   fraud scoring          soak, live                never trades          confidence

Defense-only fraud detection for cross-merchant identity reuse - catches the same UPI ID, phone, or device fingerprint recycled across merchants, with explainable verdicts and errors costed in rupees. Recommends, never auto-blocks.

• Identity link graph (fan-out + 0.6^hops taint) → deterministic 7-feature scorer → ALLOW / REVIEW / BLOCK_REC with an evidence bundle; the LLM (Bedrock) writes the analyst narrative only - it never scores
• Measured (held-out, seed 42): precision 0.833 · recall 0.882 · F1 0.857 · 2/2 rings caught · 0 silent frauds · +₹38,665 net / 1,000 events · p50 2.3 ms · total LLM spend <$0.10
• Honest disclosure: a GBDT baseline edges the rules on F1 (0.909 vs 0.857) - reported, not buried; the deterministic scorer keeps the audit contract a money-adjacent system needs
• Hardening: champion/challenger shadow scoring · Prometheus · Postgres/SQLite parity · 90% coverage gate, strict mypy, CI-gated Docker
• Stack: Python · FastAPI · Next.js 15 · AWS Bedrock · DynamoDB · PostgreSQL

Family members forward any suspicious message or payment request; a Strands agent team investigates it like a professional fraud analyst and escalates only genuine decisions to the family guardian, with a hash-chained evidence bundle.

• Strands agent loop over closed tools (link reputation · brand-claim adjudication · UPI grammar · HMAC-keyed threat graph) - the agent chooses which check to run; code composes the verdict. No tool accepts message content or returns a verdict, so the model can't launder untrusted text into evidence. Any loop failure falls back to a deterministic sweep and says so
• Measured (staging eval, real Nova Micro leg): precision 1.00 · recall 1.00 (95% CI [0.9887, 1.0]) · false-gate rate 0.0% - root-caused down from 30.6% pre-calibration and re-verified · $0.00026 / case · live soak on real households
• Graduated silence law (SILENT_KILL → AGENT_SCREEN → BADGED_RING → PASS): settled scams never page a human; degraded cases always stay visible. PII scrubbing proven by CI canary strings
• Stack: Python · Strands Agents SDK · AWS Lambda · Bedrock (Nova Micro) · DynamoDB · EventBridge · Next.js

Candidate quant trading signals investigated like fraud cases - statistical probes as tools, verdicts with receipts, silence unless a signal deserves a researcher's hour. Recommends, never trades.

• Static lint + investigator agent + 4 sandboxed verification probes (timestamp alignment · label permutation · regime subsample · turnover/cost sanity) → REJECT_SPURIOUS / NEEDS_REVIEW / PROMISING, each with numeric receipts; thresholds live in code, never in the model
• Measured (60 seeded cases, 6 strata, 48 dev / 12 sealed hold-out): spurious catch 0.925 vs lint baseline 0.475 (+0.450) · prose-hidden lookahead 0.0 → 1.0 · false-reject 0.0 · McNemar p = 0.00029 · byte-identical repro asserted in CI
• Honest changelog: a bare-prompt agent hit 1.0 catch but false-rejected 0.875 of sound signals - the four probes are the contribution that restored false-reject to 0.0
• Stack: Python · FastAPI · Typer CLI · HTMX/Tailwind (zero Node build) · sandboxed subprocess probes

A five-signal verification engine that checks LLM outputs against source documents - catching numerical transpositions, negation flips, and self-contradictions that embedding-only systems miss.

• Amazon Bedrock Titan Embeddings V2 (1024-dim) → cosine relevance + entity-contradiction engine + intra-response consistency → VERIFIED / UNCERTAIN / UNSUPPORTED, with Platt-scaled calibrated confidence (a real posterior, not a rescaled cosine distance)
• Measured (300-case adversarial benchmark): precision 95.33% · accuracy 90.33% · F1 90.79% - superiority over a cosine-only baseline proven by McNemar's test (p < 0.05)
• Serverless AWS (Lambda + API Gateway + DynamoDB) · Python + TypeScript SDKs · Next.js dashboard
• Top 50 Global Finalist - AWS 10,000 AIdeas Competition
• Stack: Python · Next.js · AWS Lambda · Bedrock

Autonomous honeypot that engages scammers in multi-turn conversation, extracts financial intelligence, and classifies scam types in real time.

• 5 adaptive LLM personas (LLaMA 3.3-70b via Groq) · 8 scam-type classifications · 85–90% intelligence extraction accuracy (phones, UPI IDs, bank accounts, phishing URLs) · Next.js dashboard · MongoDB Atlas
• Grand Finalist - India AI Impact Buildathon 2026 (Top 2% of 40,000+), presented at Bharat Mandapam, New Delhi
• Stack: Python · FastAPI · Next.js · Groq · MongoDB

Full-stack AI SaaS marketplace (Product Hunt + G2 + Gartner) for Series A–D startups - AI Trust Score, sentiment analysis, upvotes, enterprise pilot marketplace.

• Role-based access (Founder, Buyer, Admin) · 50+ REST endpoints · 7,000+ lines, solo-built and fully deployed
• Stack: Next.js 14 · FastAPI · PostgreSQL · Supabase · Clerk Auth

The part I'd want a senior reviewer to read.

• Models propose, code decides. Scoring and verdicts are deterministic and seed-reproducible; LLMs write narratives, never verdicts or money-movement decisions.
• Honest baselines. I publish where a GBDT beats my rules (Sentinel), where a bare-prompt agent false-rejects (SignalGate), and where calibration started broken (Gatehouse) - with the artifacts committed beside the fix.
• Byte-identical repro. Every claimed number regenerates from one make command with fixed seeds, asserted in CI.
• Degrade explicitly. Store down → spool-and-503 · LLM down → SKIPPED banner · probe timeout → disclosed. Never a silent pass, never autonomous blocking.
• Cost-logged AI. Per-case spend metered with breaker caps (Sentinel <$0.10 total build · Gatehouse $0.00026/case).

Languages & Systems

Agentic AI & LLM Engineering

Backend & Distributed APIs

Cloud, Infrastructure & Databases

Frontend & Mobile

Publication Title	Venue	Status	Citations / Links
Comprehensive Deepfake Detection: A Comparative Study on Image and Video Forgery	IEEE PuneCon 2025	Published
Advanced Deepfake Image Detection: A Robust Framework using InceptionV3 and Xception	i-manager's Journal on Information Technology (JIP)	Published