{{COMPANY_LEGAL_NAME}}
Top-10 Indian pharma · 12,000 employees · multi-site
{{COMPANY_LEGAL_NAME}}A Top-10 Indian pharma manufacturer reduced deviation-investigation drafting time by 60% with OwnAI.
§2 · Executive summary
Three numbers, three audiences.
Operational
−60%
Deviation-investigation drafting time
From baseline 7.2 hrs/investigation to 2.9 hrs. Measured across 80 investigations in Q1 post-go-live.
Compliance
0
Data-residency exceptions in last RBI & CDSCO inspection
All inference and audit logs on-prem. Auditor walked the system live; no follow-up CAPAs raised on AI-system data flow.
Financial
7mo
Break-even vs prior tooling
Total deal: setup + L40S + 12 mo AMC. Recouped via deviation-cycle compression and retired SaaS seats.
§3 · The challenge
Deviation backlog, audit exposure, and a per-seat AI bill that wouldn't stop.
The QA team at {{COMPANY_LEGAL_NAME}}this manufacturer was carrying a chronic 14-day rolling backlog of deviation investigations — averaging 7.2 engineer-hours per investigation, driven mostly by drafting time, not analysis. CAPAs were thorough but slow.
Two pilots with commercial AI suites had stalled. The cloud-LLM option produced credible first drafts but couldn't be deployed because routing batch-record extracts through US-hosted inference broke the customer's data-residency posture under DPDPA 2023 §8 and the team's read of Schedule M (2024). The on-prem RAG-only alternative gave citations but no useful drafting.
Meanwhile, per-seat AI tooling for 32 users was tracking to ₹38 L over five years — without solving the residency problem or producing an audit-grade trail. Procurement called for a halt.
"We needed an assistant that could draft a CAPA in our voice, on our SOPs, on our hardware. The cloud vendors couldn't say yes without breaking our DPDPA story." — Head of Quality · name withheld under NDA (real identity available on request, after mutual NDA)
§4 · The solution
An on-prem Pharma GMP pack on the customer's L40S server.
Hardware tier
L40S 48 GB · 1U Server on customer rackBase model
Qwen 3 32B · Apache 2.0 · FP16Primary use cases
Deviation-investigation drafting · SOP Q&A · Batch record anomaly flagging
Compliance artefacts
IQ / OQ / PQ shipped · ALCOA+ matrix · Langfuse log → customer SIEM
deployment.svg — on customer LAN
Timeline · pilot → production
Sep 2025
Pilot kickoff
SOW signed · eval rubric countersigned · 1,200 SOPs and 240 historical deviations transferred under NDA.
Oct 2025
Fine-tune + eval
LoRA adapter trained on RunPod A100. Eval rubric pass-bar met on 8 of 8 tasks. Cloud data destroyed within 7 days; CoD signed.
Nov 2025
Production deploy
L40S server installed on customer rack. Keycloak federated to customer IdP. Langfuse webhook wired to customer SIEM.
Dec 2025
Go-live · UAT pass
32 QA / Regulatory Affairs users onboarded. PQ certificate signed. AMC start date locked.
§5 · The results
Before vs after — measured against the eval rubric we signed at kickoff.
| Metric | Before | After | Δ |
|---|---|---|---|
| Deviation drafting time | 7.2 hrs | 2.9 hrs | −60% |
| Investigation backlog | 14 days | 4 days | −71% |
| SOP Q&A accuracy (rubric) | — | 93% | ≥ 90% bar |
| Audit-prep hours / inspection | 120 hrs | 62 hrs | −48% |
| 5-yr AI tooling spend | ₹38 L | ₹21 L | −45% |
| Data-residency exceptions | 3 (cloud pilots) | 0 | cleared |
Drafting hours per investigation
hrs · 80 investigations · Q1 2026
Before · cloud-pilot baseline
After · OwnAI Q1 2026
The win isn't just speed. It's that my QA engineer can show the auditor exactly which SOP and which version the draft cited — and the auditor doesn't have to leave the building to verify it.
§6 · Technical details
For the architect who'll defend this in an audit.
Architecture, training data, eval scores For the CTO / VP Eng
Base model
Qwen 3 32B at FP16 precision Apache 2.0
Adapter
LoRA · rank 16 · 168 MB · version
v1.2-pharma-deviation. Trained on RunPod A100 80 GB; cloud volume destroyed within 7 days of delivery (CoD signed).
Hardware
L40S 48 GB · 1U server on customer rack. Sustained 62–78 tok/s at batch 8, 8K context. Air-gapped from public internet.
Training data volume
1,200 SOPs · 480 batch records · 240 historical deviations · 60 CAPA reports. All on-prem; never sent to a third-party labelling service.
RAG corpus
Qdrant · 18,400 chunks · BGE-M3 embeddings · per-RBAC payload filtering tied to Keycloak groups.
Stack
vLLM · LiteLLM · Keycloak · Langfuse · Prometheus/Grafana · Caddy · Restic
Audit pipeline
Langfuse → webhook → customer SIEM (Splunk). Hash-anchored entries; NTP-disciplined; full ALCOA+ map at /pharma#alcoa.
Evaluation scores · pilot SOW rubric
| Task | Pass bar | Result |
|---|---|---|
| Deviation draft · factual correctness | ≥ 90% | 97% |
| SOP Q&A · citation accuracy | ≥ 90% | 93% |
| Batch record anomaly recall | ≥ 85% | 89% |
| Change control impact draft | ≥ 85% | 88% |
| Hallucination rate (lower is better) | ≤ 2.0% | 0.6% |
Same shape of problem?
Want similar results? Book a discovery call.
A 30-minute call with the founder. Bring three of your hardest deviation reports — anonymised is fine — and we'll walk through how the drafting assistant would have handled them.