Pragyan
AI-Enabled Cadre Competency Auditing, Automated Skill Gap Diagnostics, and Capacity Building for India's Official Statistical System (MoSPI / SIH26101).
The Problem & Context
At Smart India Hackathon 2026, I led a 6-member engineering team (Byte Force) tackling problem statement SIH26101: building an intelligent competency auditing system for the Ministry of Statistics and Programme Implementation (MoSPI).
Official statistical personnel (Indian Statistical Service, Subordinate Statistical Service, and State DES) manage massive national operations including census enumeration, PLFS labour rounds, GDP compilation (SNA 2008), and CPI price indices. However, their Annual Performance Appraisal Report (APAR) training recommendations were subjective, siloed, and disconnected from the 800+ courses on the government's iGOT Karmayogi portal.
Our mandate: replace subjective guesswork with automated AI skill-gap diagnostics, dynamic assessment generation from official training PDFs, and direct course matching aligned with the Capacity Building Commission (CBC).
System Architecture
I architected the platform as a full-stack Next.js 14 App Router application with zero-downtime AI failover:
1. Triple-AI Dynamic Router (Groq 70B → Mistral → Gemini)
Instead of relying on a single brittle AI endpoint, I built an abstraction layer that intercepts user prompts. It prioritizes Groq (Llama-3.3-70b-versatile) for sub-second streaming inference. If rate limits or network degradation occur, it seamlessly falls back to Mistral Small, and finally to Google Gemini Pro. Token streaming to the client never disconnects.
2. MoSPI 4-Domain Competency Matrix
We engineered a quantitative scoring framework across 4 mandatory domains: Statistical (Sampling, National Accounts, Estimation Theory), Technical (Python/R, SQL, GIS Spatial Analysis), Digital Governance (DPDP Act 2023, NDGFP), and Behavioral (Leadership, Field Coordination). Each cadre profile (ISS Director vs SSS Investigator) is benchmarked with custom baseline-to-target deltas.
3. Non-Hallucinatory Policy Ingestion Engine
Built a server-side document parser using pdf-parse and mammoth. Officers can upload raw ministry manuals, NSS survey guidelines, or gazette PDFs. The engine chunks text, extracts key statistical propositions, and forces deterministic JSON output to create non-hallucinatory diagnostic tests with official citations.
4. Executive Telemetry & APAR Dossier Generator
Using Recharts, I developed live radar charts comparing cadre skill overlap, division-wise capability distributions (NSO, CSO, NSSO, NAD), and an APAR-compliant printable audit certificate that officers can directly attach to their annual appraisal submissions.
The Hardest Engineering Challenges
1. Eliminating Hallucinations in Statistical Assessments
Standard LLM prompts hallucinate statistical formulas (e.g. ratio estimators or FISIM banking calculations). I solved this by implementing few-shot JSON schema enforcement combined with domain grounding: questions must strictly cite an excerpted paragraph from the uploaded manual, and distractor options must reflect authentic survey operational errors.
2. Universal Neo-Brutalist Responsive UX
Government portals are notoriously clunky and desktop-locked. I designed Pragyan with a high-contrast Neo-Brutalist design language (deep black #000000, cherry red #e11d48, crisp monospace badges). On mobile viewports, I converted the fixed desktop sidebar into an off-canvas drawer with backdrop blur, enforced 44px touch targets, and engineered horizontal touch-scrolling for 6-column personnel matrices.
3. Production Resilience & Zero-Config Demo Mode
During hackathon evaluation, evaluators often test portals without configuring API keys. I built dual-state operational resilience: when live API keys are provided, the portal uses live Llama 3.3 70B inference; when keys are omitted, the system falls back to simulated MoSPI intelligence with preloaded assessment packs (PLFS, SNA 2008, DPDP 2023) ensuring 100% uptime.
Takeaways & Leadership
Leading Byte Force at Smart India Hackathon 2026 was an intense crash course in rapid systems architecture, delegation, and endurance. While hackathon judging outcomes are subjective, taking the prototype and refining it into an open-source, production-ready system with verified Next.js builds, mobile parity, and Google Search Console indexing demonstrated the difference between a prototype and an enduring product.