نبذة عني
AI engineer building production agentic systems for the enterprise — multi-agent orchestration, RAG/GraphRAG grounded in company data, and natural-language-to-action pipelines, including a text-to-SQL agent at ~98% accur…
AI engineer building production agentic systems for the enterprise — multi-agent orchestration, RAG/GraphRAG grounded in company data, and natural-language-to-action pipelines, including a text-to-SQL agent at ~98% accuracy over a 97-table schema. AI Engineering Intern at FutureCraft and SP Jain; prior SAP HCM consulting. Trusted AI by design: grounded, schema-verified, approval-gated, audit-trailed.
الخبرة
AI ENGINEERING INTERN
<p>Integrated Google Gemini (Flash) for real-time lecture recap.</p><p> Architected an additive, feature-flag-gated AI co-teacher layer across 4 production Angular/Node.js apps for a live university-classroom platform.</p><p> Built grounded-response guarantee: fully inert when disabled, honest “unavailable” over fabricated content.</p><p> Implemented activity generation, student Q&A, and auto-notes, grounded in live speech-to-text.</p><p> Built a real-only response envelope (retry, prompt caching, tolerant-JSON parsing) with TDD contract tests.</p><p> Built an on-device MediaPipe attention signal (fully in-browser, score-only, privacy-by-construction — no video or facial data leaves the device).</p><p> Fused attention with answer-correctness and participation into a deterministic per-seat engagement engine.</p><p> Implemented P2P live video (WebRTC publisher → per-seat subscriber) over a socket.io signaling relay, delivering 6 interactive activity types.</p><p> Led an adversarial security review surfacing a critical WebRTC authorization gap plus cost-DoS and prompt-injection risks.</p><p> Migrated the classroom media plane from Intel OWT to a self-hosted LiveKit SFU across five repositories.</p><p> Caught remote audio dead in two apps after three green reviews because every check asserted on video.</p><p> Audited 99 socket handlers — only two checked identity, and a student token reproduced cross-class roster disclosure live.</p><p> Built and differential-tested a 127-arm authorization layer that runs in shadow mode.</p>
AI ENGINEERING INTERN
<p>Deployed the Prajna Knowledge System end-to-end on AWS.</p><p> Led the text-to-SQL pipeline for GYCO, an enterprise BI platform.</p><p> Enabled non-technical users to query operational and financial data in plain English.</p><p> Re-architected a failing multi-agent pipeline into SQLAgent v2 (dense LLM + vLLM speculative decoding) — ~98% accuracy on a 50-query benchmark over a 97-table schema.</p><p> Engineered schema-grounded generation with forced verification — every query validated against schema contracts pre-execution.</p><p> Implemented a query/cell/user-level audit trail, mirroring regulated-data dual-validation governance.</p><p> Delivered (ECR/EKS/EC2) in two days ahead of a customer-facing stakeholder demo: IAM, networking, container orchestration and secrets management.</p><p> On a VARA-regulated asset-tokenization engagement, identified a document-classification compliance gap and audit-trail issues before client deployment.</p><p> Authored the architecture docs and onboarding material the team builds from.</p><p> Built an OpenAI-compatible mock with 10 frozen scenarios so the team and CI could build without a GPU.</p><p> Served Qwen3.8-27B GPTQ 4-bit on vLLM.</p><p> Engineered a five-layer hallucination stack in which the narrator receives placeholder tokens instead of digits, so any figure the model writes is provably invented.</p><p> Built a 65-case adversarial suite that caught client IDs leaking through summary statistics.</p><p> Owned the LLM lane on GYCO Finance, a four-person portfolio.</p>