I'm an engineer who likes building LLM systems that don't fall apart once real users touch them. Most of my work sits between production retrieval pipelines, agentic workflows, and the dependable backend infrastructure required to keep them deterministic.
Currently building AI products at Monsoonfish, with prior stints across two other product teams. When I'm off the clock, I'm usually deep-diving into AI architectures, building things that actually solve problems rather than just look good in a demo, or grinding LeetCode (Knight).
Outside of code: football and the gym.
An AI Engineer / Backend Developer.
pune, india · trinity academy · be cs 2027
CHOOSE YOUR PATH
Production-grade AI systems, multi-agent frameworks, and retrieval pipelines.
Photograph a prescription, get structured records and a chatbot that answers only from your own history.
AI reads English prescriptions with Gemini, and Hindi or Marathi ones with a Sarvam OCR + Gemini hybrid. A RAG chatbot then answers questions from your records only, while allergy and drug-interaction checks warn about unsafe medicines.
Explainable video forensics that tells real footage from deepfakes and AI-generated video.
Instead of one black-box score, a panel of specialist agents (face-swap, motion, audio, C2PA provenance, vision LLM) examines each video in parallel. A LangGraph consensus engine returns AUTHENTIC, MANIPULATED or UNCERTAIN, with a reason for every flagged frame.
GitHub retrieval-intelligence workbench.
Shows which retrieved items actually made it into an answer and which were dead weight, using hybrid vector + weighted-graph retrieval.
A new build is being forged at the camp.
Check back soon. This slot unlocks when the next project ships.
Turns a photo of a prescription into structured data, then lets you ask questions about your own medical history.
Take a photo of a prescription and the AI reads the medicines, dosages, symptoms, vitals and allergies. You review and edit the result before saving it. After that you can ask a chatbot things like "what was my last BP?" or "which antibiotics have I taken?", and it answers only from your own records.
Tested on 50 synthetic prescriptions (English, Hindi, Marathi) and 70 questions over 49 synthetic records, because real medical data can't be used. The right record is retrieved 98.4% of the time, 9 of 9 out-of-scope questions were refused, no records leaked between patients, and the median answer takes about 1.1s. The hybrid is 1.6x slower than Gemini only (3.8s vs 2.3s), both engines scored 100% on English, and its Indic lead is 1 medicine out of 56. The images use clean fonts, so read these as an upper bound, not real-world accuracy.
Decides whether a video is authentic, a face-swap deepfake, or fully AI-generated (Sora, Veo, Kling, Wan).
Most detectors give you a single opaque score. VeriFrame works like a panel of experts looking at the same video: each agent examines a different kind of evidence, a LangGraph state graph coordinates them, and a consensus engine combines their findings into one verdict with per-frame, human-readable explanations streamed live to the UI.
By generator: Sora 2 7/7, Veo 3 6/6, Kling 4/6, Wan 2 3/6. Real HD footage: 0 of 12 falsely flagged.
Measured on a reproducible 50-video benchmark: 25 AI clips from Sora 2, Veo 3, Kling and Wan 2, and 25 real clips (HD and low-res). It is a small sample (95% confidence interval for accuracy: 76-94%), so treat it as an internal benchmark. Videos the system called UNCERTAIN count as "not flagged"; counted strictly as wrong, accuracy is 80%. C2PA Content Credentials did a lot of the work: all 13 Sora 2 and Veo 3 clips carried a signed AI label, which is 13 of the 20 detections, and re-encoding or social-media upload strips them. The Gemini/Groq reasoner alone flags 13 of the 25 AI clips and none of the real ones. Face-swap datasets were not part of the benchmark, and latency was measured on a laptop, so Render's free CPU will be slower.
Retrieval-intelligence workbench analyzing answer-overlap vs dead weight in GitHub retrieval.
Asks "of the items retrieved for an answer, which actually appeared in that answer — and which were dead weight?" Features hybrid (vector + weighted-graph) retrieval with intent routing, evidence confidence tiers, and a client-side "what-if" fusion slider.
My journey across AI product teams, production systems, and engineering leadership.
My technical toolkit across production AI, multi-agent frameworks, and dependable backends.
A glimpse into my open-source contributions and pinned projects.
Pinned Repositories
Medical-records intelligence platform with a hybrid OCR pipeline and grounded Q&A with allergy safety guardrails.
Multi-agent video forensics platform that detects AI deepfakes with explainable per-frame verdicts.
GitHub retrieval-intelligence workbench showing which retrieved items made it into an answer and which were dead weight.
Competitive programming, student leadership, and off-the-clock interests.
Grinding DSA to Knight rank on LeetCode (@sahil_2610), solving over 600+ problems across graphs, trees, and dynamic programming.
Serving as Event Coordinator for ACES at Trinity Academy of Engineering, organizing department-wide coding contests and hands-on workshops for 100+ students.
Researching retrieval architectures, AI automation, and LLM inference & evaluation: prototyping is easy, sustaining it under real-user load is the hard part.
Outside of code: passionate about football, dedicated to the gym, and discussing movies & technology.
Milestones and achievements unlocked along the journey.
Earned the Knight Rank on LeetCode with 600+ solved problems across algorithmic paradigms and multiple coding platforms, ranking in top competitive percentiles.
Maintained an academic CGPA of 8.6 / 10 throughout the B.E. Computer Science program at Trinity Academy of Engineering, SPPU.
Led MyUXScore from prototype to production-ready deployment; now adopted as an internal tool used daily by 50+ employees across the company.
Won the college-level hackathon TriniHack by building CampusKart, an all-in-one engineering equipment resale platform for students, featuring a real-time bidding system.
Let's build reliable AI systems together. Reach out!
I'm an engineer who likes building LLM systems that don't fall apart once real users touch them. Most of my work sits between production retrieval pipelines, agentic workflows, and the dependable backend infrastructure required to keep them deterministic.
Currently building AI products at Monsoonfish, with prior stints across two other product teams. When I'm off the clock, I'm usually deep-diving into AI architectures, building things that actually solve problems rather than just look good in a demo, or grinding LeetCode (Knight).
Outside of code: football and the gym.
Curious experiments, visual rolls, and interactive surprises.
Click to explore the interactive scroll animation. It goes further than you think.
Classified / In Development.