Capstone Library — What Kompas Students Ship

Problem

ASHA workers record patient interactions in code-switched Hindi and English. Off-the-shelf Whisper-class models hallucinate medical entities and miss Indic phonology, making the transcripts unreliable for clinical follow-up.

Target outcome

A domain-adapted Whisper-derivative reaching >90% medical-term accuracy on a held-out 1,200-utterance evaluation set, with a privacy-by-design audit. Deployed at three primary health centres for a four-week field pilot.

Problem

Court researchers need to retrieve precedent across 1.2M judgments without hallucinated citations — the failure mode that disqualified prior LLM trials.

Target outcome

A hybrid retrieval pipeline (dense + BM25) with a citation-verification step that refuses to answer when retrieval confidence is below threshold. Zero hallucinated citations across a 500-query evaluation set, defended in a registrar-style audit.

Problem

A live B2B SaaS spending ~₹14 lakh/month on inference; mixed routing across closed-source frontier models and open-weights deployments without a defensible cost model.

Target outcome

A routing layer (cost × latency × quality) that cuts inference spend by ~35–40% with no measurable quality regression on the customer-facing benchmark. Includes an evaluation harness so the team can defend future routing decisions.

Problem

An IoT device with sub-2 GB RAM needs on-device language understanding for offline-first usage in low-connectivity Indian field deployments. Frontier 7B-parameter models are far too large.

Target outcome

A quantised + LoRA-fine-tuned sub-1B-parameter model running at <500 ms/inference on the target hardware, with documented quality-regression bounds against the un-quantised baseline. Open-source quantisation pipeline + reproducible benchmark report.

Problem

Existing NLP benchmarks under-test code-switched Indic input. Without rigorous evals, no Indian team can defensibly choose a model for production deployment — the choice is currently driven by vibes.

Target outcome

A reusable evaluation harness covering 5+ task types (sentiment, NER, summarisation, retrieval, dialogue), a public leaderboard with ≥10 baseline runs, and a methodology paper submittable to an Indic-NLP workshop.

Problem

Digitising 19th-century Devanagari and Bengali archives requires OCR that handles handwritten scripts and degraded paper. Current commercial OCR is English-first and degrades sharply on Indic glyphs and archival print runs.

Target outcome

A four-stage pipeline (image-quality → OCR → post-OCR correction → layout reconstruction) achieving >85% character accuracy on a held-out 2,000-page corpus. Pipeline released open-source; the processed corpus returned to the partner archive.

Problem

A mid-tier non-bank lender wants to deploy AI in underwriting and collections without tripping RBI guidance or DPDP exposure.

Target outcome

A 12-page AI-readiness audit with a sequenced 18-month roadmap, a vendor-evaluation framework, and a board-ready risk-and-compliance memo. Defended to the lender's CXO panel.

Problem

A 200-employee manufacturing exporter has stagnated lead generation. The existing CRM and content stack predate the LLM era and lack usable customer data.

Target outcome

A reconstituted GTM stack with LLM-driven segmentation and outbound personalisation. Pilot quarter shows a 2× or better rise in qualified leads at half the prior tool cost. Includes a measurement plan the marketing lead can run after the cohort leaves.

Problem

An enterprise IT-services CIO faces six competing AI vendor proposals; no consistent way to compare claims, total cost of ownership, or DPDP/EU AI Act exposure.

Target outcome

An open-source vendor-scorecard framework adopted internally by the IT team and shared with peer companies. Drives a 5×+ reduction in vendor-pitch evaluation time, with documented criteria the procurement function can defend.

Problem

A six-hospital tier-2 chain is considering AI for scheduling, billing, and discharge summaries. The board has no clarity on where AI helps versus where it creates risk under DPDP and healthcare regulation.

Target outcome

A workflow audit, a sequenced 12-month rollout plan covering three high-impact areas, a board-ready risk-and-compliance memo, and a change-management playbook. Defended to the hospital's executive team.

Problem

A 300-employee SaaS company uses AI résumé-screening; the HR head suspects bias but lacks internal capability to audit. Legal exposure under the EU AI Act and India's emerging frameworks is unmodelled.

Target outcome

A documented bias audit (disaggregated metrics + adversarial testing), a vendor-replacement recommendation, and a hiring-AI governance policy adopted by the leadership team. Includes a quarterly re-audit cadence the HR function will run.

Problem

A growing D2C brand's CEO wants to deploy agent-augmented customer support but cannot defensibly model the expected P&L impact for the board.

Target outcome

A model quantifying expected change in cost-per-ticket, time-to-resolution, NPS, and gross margin under three deployment scenarios. Risk-adjusted with sensitivity analysis, defended to the CFO. Includes a 12-month phased rollout plan and a kill-switch criterion.

Problem

A partner university's existing faculty across CSE, Business, Law, and Liberal Arts need AI fluency to support — and eventually co-deliver — Kompas tracks themselves.

Target outcome

A 12-week cohort producing ~20 faculty members fluent enough to co-teach Track D modules and contribute to curriculum refresh cycles. A subset signs on as long-term Kompas-trained co-faculty.

Problem

A partner research group with a strong domain background but no prior AI-publication track wants to ship a peer-reviewed paper within 18 months without abandoning their domain.

Target outcome

A joint paper accepted to a workshop at a top-tier ML or domain conference within 14 months of lab launch. First DST/ICMR grant submission filed with Kompas grant-writing support; pending review.

Problem

A partner department's AI-related curriculum is 2–3 years out of date. The chair needs a defensible refresh process the department can sustain on a 9-month cadence after the cohort leaves.

Target outcome

A 6-week curriculum-refresh sprint: four modules rewritten, 6+ industry signatories on the curriculum review, and a 9-month-cadence refresh process formally adopted by the academic council. Templates and rubrics handed off so the next sprint runs without Kompas.

Problem

Non-CS faculty teaching AI (law, journalism, design, business, liberal arts) struggle to avoid the two failure modes — AI-for-poets and B.Tech-lite. No accessible pattern library exists for the patterns that actually work.

Target outcome

A curated library of 30 teachable patterns with worked examples across disciplines, plus a faculty training cohort that has internalised them. Library released openly with an annual refresh schedule.

Problem

A research-strong partner department has never accessed AI-specific national grants. Existing grant infrastructure doesn't fit AI-proposal shapes, and faculty under-estimate the documentation effort.

Target outcome

A grant-funnel process — opportunity scan → proposal scaffold → review cycle → submission — institutionalised in the partner department. Three submissions in cycle 1; at least one funded within 12 months.

Problem

The partner university wants to translate Kompas-track student work into incubation outcomes, but the existing innovation-cell process is generic, slow, and AI-illiterate.

Target outcome

An AI-track-aligned innovation cell with an explicit student-startup pathway. Cohort 1 incubates three student ventures; one raises external pre-seed within 12 months. Process documentation handed off so cell can run cohort 2 independently.