PhysicsWallah · AI Innovation · IOI (Institute of Innovation)
Tools, Technologies, Frameworks & Methodologies
Internship work — two production systems for the PW IOI admissions & onboarding teams.
Period: June – August 2026 · Systems: PW Intelligence (admissions analytics + forecasting) & IOI Onboarding Portal (fee reconciliation) · Both live and in daily use.
01Programming Languages
| Python 3.12 | Primary language for both systems — data ingestion, the AI agent, the forecasting engine, backend APIs, and automation. |
| SQL | The analytical core — hand-written and AI-generated queries across two dialects (DuckDB and PostgreSQL). ~205 tables/views modelled. |
| JavaScript (vanilla) | Front-end interactivity and charts with no build step or framework — deliberately lightweight. |
| HTML / CSS | Responsive, theme-aware UI for the dashboards and chat interface. |
02AI / ML Models & Techniques
| Google Gemini 2.5 | LLM family (flash / pro / flash-lite) powering natural-language → SQL translation, reasoning, and narration of results. |
| Agentic ReAct architecture | A reason → act → observe loop with read-only database tools: the agent investigates, cross-checks sources, flags mismatches, then answers — not a single prompt. |
| Retrieval-Augmented Generation | A curated store of verified question→SQL "golden" examples injected as few-shot context — the single biggest accuracy lever in NL→SQL. |
| Deterministic forecasting | A 3-method ensemble (pipeline-inventory × conversion-lag curves, seasonal-pace, momentum) — structure over statistics, so every number is explainable and reproducible. |
| Grounding & guardrails | A numeric grounding guard (answers must tie to query results), a pinned metric registry (same question → same number), and prompt-injection hardening. |
| Evaluation | LLM-as-judge red-team harness (100+ graded queries) plus a forecast back-testing harness that replays the engine against history. |
03Data Tools & Storage
| DuckDB | In-process analytical database (~205 tables, 550k+ rows) — the query engine behind PW Intelligence. |
| PostgreSQL (Supabase) | Cloud operational database (ap-south-1) for the onboarding portal and the analytics clone. |
| Google Sheets API (gspread) | Live ingestion of 9+ source sheets, header-drift–resilient, on a scheduled sync. |
| Frappe / ERP REST API | Live student, fee, and hostel data pulled from PW's ERP (whole cohort in 3 bulk requests). |
| sqlglot | SQL parsing & validation — the deterministic guard that keeps the AI's queries read-only and safe. |
| Identity reconciliation | Cross-source matching (ERP ↔ Medhavi/MSU ↔ loan) by ID → email → mobile to unify each student's payment status. |
04Backend, Web & Design Frameworks
| FastAPI + Uvicorn | Async (ASGI) web framework for both systems' APIs and pages. |
| Jinja2 + Tailwind CSS | Server-rendered UI for the onboarding portal (no Node build). |
| Chart.js | Interactive dashboards, funnels, and forecast trajectory charts. |
| Server-Sent Events | Live streaming of the agent's reasoning steps to the chat UI. |
| Auth & security | Session-cookie auth (itsdangerous), PBKDF2 password hashing, role-based access control (4 roles), append-only audit logs. |
05Cloud, DevOps & Deployment
| GitHub + GitHub Actions | Version control and CI/CD, incl. a scheduled Action that rebuilds the analytical dataset and publishes it to cloud storage. |
| Vercel | Serverless hosting for both systems (region-co-located with the database for 20–30× faster responses). |
| Supabase | Managed Postgres + object Storage + Cron (scheduled background jobs). |
| Cloudflare | Tunnels (secure VM exposure, no open ports), Pages (static hosting), and DNS. |
| pm2 | Process management for the VM-hosted services (web, data-sync, automations, tunnels). |
06Methodologies & Engineering Practices
| Agentic AI design | Tool-using LLM agents with strict read-only boundaries, self-healing on errors, and human-in-the-loop feedback that writes back correction rules. |
| Reproducibility | Temperature-0 generation + deterministic answer caching — the same question always returns the same, audited number (critical for leadership trust). |
| Test & validation-driven | pytest unit tests, a red-team eval harness, and forecast back-testing (measured ~1% day-of admissions error) before any number reaches leadership. |
| Data governance | Real student PII handled under DPDP-aware practices — read-only sources, no raw rows in logs/output, secrets isolated and gitignored. |
| Graceful degradation | Every external source (sheets, ERP, LLM) degrades to last-good data instead of failing, with alerting. |
| Incremental migration | Zero-downtime SQLite→Postgres and VM→cloud migrations via thin compatibility shims and env-gated, backwards-compatible code. |
| AI-assisted development | Built using agentic coding tools (Claude Code) with a "ship the simplest thing that works" discipline and rigorous review. |