Career
Where I've
built things.
DATA / ML ENGINEER
- Building the academic co-authorship edge pipeline (pubnet) — a new graph edge type joining co-worker, co-student, and co-athlete; ingests 250M+ entities from OpenAlex and OAG into a bronze-normalize-load pipeline with deduplication via canonical_work_id and versioned edge policy scoring
- Designed a multi-stage canonicalization framework for matching scraped entities to ~900M LinkedIn person nodes in Postgres — ground-truth-first methodology using name, university, grad year, company, and hometown signals; tiered candidate filtering to protect production DB under live traffic
- Conducted data source feasibility evaluations (Vieu Standard format) across OpenAlex and OAG — produced canon rate reconciliation, joint Tier-1 edge resolution estimates (~3.5–3.8% indicative), and a 74% discard-rate analysis informing pipeline design tradeoffs
- Working in a pre-computed graph architecture (Postgres + OpenSearch) powering B2B relationship intelligence — Lambda-based ingestion pipelines, event-driven edge computation, and path-finding for sales outreach across a 900M-node person graph
Pre-seed · 10-person team · SEBI-registered algo trading startup
FOUNDING AI ENGINEER — AAGMAN AI
- Architected a 5-layer neuro-symbolic trading platform (0→1) — SEBI April 2026 compliant: LLM planning (Mastra AI) + deterministic OPA/Rego policy enforcement, guaranteeing zero hallucination-induced regulatory breach
- Designed a JSON DSL compiler as a safe LLM output format — AI selects building blocks, interpreter executes them; eliminates arbitrary code execution and prompt injection risk in financial execution
- Built VectorBT-backed backtesting engine with golden test suites verifying exact trade counts, timestamps, and metrics across all executions — deterministic by design
- Engineered vectorized batch screener processing 500+ instruments per run at >10× sequential throughput; real-time ClickHouse + TimescaleDB ingestion pipeline with multi-broker adapters (Zerodha, Groww, Angel One, Upstox)
- Designed HNSW-indexed RAG memory (pgvector) with hybrid BM25 + dense retrieval; intent router (distilled Llama-3 8B + Redis semantic cache) reduces full LLM calls by >90%, cutting per-query cost from $0.10 → $0.01
- Mentored 2 engineers on deterministic execution design, OPA policy authoring, and CI/golden-test methodology
DATA SCIENTIST — RANKING, PERSONALIZATION & PRICING
- N1 user-level personalization — replaced 25-cohort system; +150% Click Recall@50 (12%→30%), +16% buyer conversion uplift, <100ms p99 serving latency — 1M+ monthly sessions; solved cold-start for zero-click users (35% of base) via implicit search/filter signals
- GBDT Default Sort V2 — dynamic de-boosting of underperforming inventory; +21% V2Bi uplift (top-5 deciles), +7% clicks per vehicle; out-of-time validated + live A/B confirmed; PSI monitoring for ongoing drift detection
- FSP pricing model — vehicle fingerprint + demand/supply + LLM-structured inspection data; ~70% predictions within ±5% of actual sale price; multi-dimensional OOT evaluation framework
- Similar Cars hybrid recommendation (70:30 collab:content) — +6% impressions, +12% SimilarCar U2Bi; resolved exploration-vs-exploitation tradeoff
- Served recommendations at <100ms p99 latency via Two-Tower retrieval model with HNSW ANN index; expanded personalised coverage from 65% → 100% of user base
- Migrated 9 DS models to GA4 with zero downtime; owned cross-geo DS for UAE, Thailand, and Australia simultaneously
- Mentored 2 junior analysts on SQL-based A/B test validation, out-of-time evaluation methodology, and ML pipeline best practices
DATA SCIENTIST · INTERN → FULL-TIME
- Logo-similarity microservice — embedding-based CV deployed on AWS S3 + Hive; 400M+ comparisons/run, 900k backlog records, SLA-bound to fortnightly Fortune 500 data deliveries; 4% false-positive reduction at millions-of-detections scale
- Contributed to Togylop — 6sense's internal NLP training library (BERT/RoBERTa multi-class, multi-label, token classification); listed as library maintainer
- Scaled B2B intent taxonomy 57 → 198 divisions (12 → 15 functions) via supervised entity classification; signals consumed by Fortune 500 ABM workflows
DATA ANALYST
- Built automated ETL pipelines and analytics dashboards for policy monitoring, consolidating multi-source departmental datasets into standardised reports for senior government officials.