Edara Sri Sai Kaushal
AI Engineer. I build production multi-agent systems that run unattended.
AI Engineer Intern (PPO)
Kookar AI
Jun 2025 — Present
IIIT Sri City
B.Tech (Hons.) CS
Nov 2022 — Jun 2026 · CGPA 9.2/10
Civic issue reporting + public accountability layer for India. Citizens photograph an issue, it gets geo‑tagged and routed to the responsible department, then tracked publicly. Includes a curated directory of 4,800+ urban local bodies, central & state grievance officers, and district‑level contacts.
Nutrition tracker built for Indian families — snap a photo of your meal and instantly see calories, macros, micronutrients, and diet‑quality insights. Visual dashboards give you daily/weekly trends, deficiency alerts, and portion guidance so you stay in control without manual entry. Multi‑user profiles let every household member track together, and an admin console helps nutritionists or caregivers oversee plans.
AI Engineer Intern (PPO)
Kookar runs households end to end with LLM agents — meal planning, grocery ordering, cook coordination and customer communication over WhatsApp.
Turns meal plans into placed grocery orders. Demand- and expiry-aware stock netting, cross-platform cart allocation, idempotent end to end.
Coordination layer making 20+ independently running agents act as one system per household.
Replaced a 30-minute scheduled operator call. Resumable, stitches multi-session context, auto-triggers downstream setup.
Multi-agent across customer and cook sides. Extracts every household detail, takes a lead from first contact to a live kitchen.
Catches agent errors before they reach the customer.
Lets RMs manage households from WhatsApp alone, without ever opening a dashboard.
Problems reach the right human in seconds instead of 22 minutes; an agent drafts the fix.
Per-agent model routing, cohort overrides, HITL review queues, plus growth and funnel analytics dashboards.
A hand-built Postgres warehouse under the dashboards: a view-migration suite plus cron scheduling covering cohort analysis, phase classification, retention, funnel and growth. On top, a reporting layer that attributes LLM cost per project from LangSmith traces — FinOps, not just charts.
The end-user surface. Expo / React Native with phone-OTP auth, Razorpay checkout, a transactions view, an auth-sync trigger, and an onboarding + scheduling rewrite on a custom calendar.
Structured-output repair and recovery policies, a provider-capability model, a fallback stack, a litellm gateway behind service tokens, and model routing lifted out of code into a live agent registry with BYOK presets.
ALB, EC2, IAM, Route53, SSL, secrets and security modules with a bootstrap and deploy script, plus OAuth login wiring. I have also decommissioned a full stack, cross-stack references included — build and teardown.
My first task there: a customer-facing voice product from nothing to a live TLS domain in four days, alone, deployment included.
Research Intern
Fog computing for BCI healthcare, mentored by Dr. Arijit Roy. Two results — one of them peer-reviewed and published.
Fog computing for BCI healthcare. Mean episode latency cut against the plain-DDQN baseline.
−61%latency · Wilcoxon p = 0.001 · n = 10 seeds
Head-selection algorithm for fog-enabled EEG healthcare IoT. Balances latency, energy and reliability across the topology.
LatencyEnergyReliability
E.S.S. Kaushal, R.K. Rath, S. S.R., A. Hazra. Proc. IEEE INDICON, 2025.
IEEE Xplore ↗ Written during the IIT Patna internship
De-novo molecule generation — a generative-chemistry model taught to write valid molecules, then rewarded into writing good ones.
Two-stage pipeline. Stage 1 trains an LSTM on 1M+ SMILES for next-token prediction. Stage 2 fine-tunes it with REINFORCE against a custom reward combining validity, novelty, fragment presence and QED. Top-p (nucleus) sampling for generation diversity, plus fragment-based control to steer toward a scaffold. An automated evaluation harness scores validity, novelty, uniqueness, diversity and QED.
Stack TensorFlow 2.19 · RDKit · Python 3.10 · LSTM · REINFORCE Repository ↗
Moves surplus food from restaurants and households to families in slums, matching every donation to the nearest delivery partner with race-safe handoffs. Express + MongoDB behind a Next.js frontend, JWT auth with bcrypt. Redis on Upstash caches the post feed with write-through invalidation on new posts; MongoDB indexes cut query time. Swagger docs for every route at /api-docs, Jest unit tests, a Dockerised backend shipped to Docker Hub and deployed on Render via GitHub Actions, frontend on Vercel.
Source Race-safe donation handoff · write-through cache invalidation Repository ↗
LSTM + Fuzzy-ART deep RL scheduler with per-prototype replay buffers, driving a discrete-event simulator under Poisson node churn. Separate fog-head and fog-node models, a baselines suite and a benchmark harness in-repo.
Stack Python · Deep RL · LSTM · Fuzzy-ART · discrete-event simulation Repository ↗
B.Tech (Hons.) Computer Science.
9.2/ 10 CGPA
Directed 10 technical clubs and ran the central and alumni portals.
1,500+ students4,200+ alumni6,500+ at Abhisarga fest10clubs
Ran workshops, hackathons and contests for the campus developer community.
2+ workshops3 hackathons12 coding contests200+students
Swallowed exceptions · a cached False with no TTL · RLS on with no policies. Every fix ships instrumentation.
Triggers · CHECK constraints · partial unique indexes. Make the bad state unrepresentable.
Traces decide, not static analysis. A path comes out once production shows nothing reaches it.
Anchored to a real broken row, proved inside a transaction I then rolled back.
Code ↔ migration ordering, stated — including the windows where the database would outrun the app.
HITL settings · playbooks · model routing — tables operators edit without a deploy.