Abhay Prakash

Winner among 120 projects as team lead; delivered a working multi-service backend under tight deadlines.

Track 3, Unicorn AI Product Building Challenge.

LLM APIs, RAG, fine-tuning, and agentic AI systems.

Built an agentic insurance-claim assessment platform combining text and visual evidence — Ingestion → Parsing → Grounding → VLM Analysis → Deterministic Rules Engine → Decision Output — with graceful degradation under provider failures.

Layered a deterministic rules engine (Supported / Contradicted / Not Enough Information) over VLM outputs to reduce hallucinations, producing auditable decisions with image-level evidence citations and adversarial/prompt-injection resistance.

Designed a real-time RESTful backend (Flask + NodeJS) serving gesture-recognition inference to a ReactJS frontend; built an NLP + CNN pipeline achieving 80% accuracy. Led a team of 6 as Full-Stack Developer in HackerWar 5.0 (SIH'24); won among 120 competing projects.

Built a 9-stage deterministic reasoning pipeline (Evidence Graph → Belief → Contradiction → Ecology → Trait → Reflection → Probability) processing 100,000 candidate profiles from a 487MB JSONL stream at a constant ~0.5MB memory footprint. Implemented an anti-gaming validator cross-checking career-timeline and skill-duration claims, flagging and rejecting 9,325 candidates (9.3%) across 46,600+ logged inconsistency events.

Designed a contradiction-detection layer scoring exaggeration severity directly into candidate rank, producing reasoning strings that trace every claim back to a specific evidence node instead of free-text LLM output. Wrote and maintained PyTest unit tests validating pipeline components against regressions.

Designed a 6-dimensional reliability feature pipeline (semantic similarity, NER consistency, numerical consistency, response stability, lexical overlap, language consistency) and integrated 5 heterogeneous datasets (BHRAM-IL, PsiloQA, HaluEval, WikiBio-GPT3, Indian-domain corpora) into a unified evaluation corpus. Implemented a Detect–Inhibit–Rechannelize remediation loop: stochastic 3-run generation at increasing temperature to surface response instability, followed by Wikipedia-RAG-grounded regeneration at lower temperature for correction.

Built the classifier training/evaluation harness (Logistic Regression, Gradient Boosting, Random Forest with stratified CV) as a reusable pipeline component.