~20 min → <4 s
Per-PDF review time
Illustrative POC/MVP reference: manual review versus local-GPU assisted processing.
Case study
Forensic drug analysis — lab report review
An illustrative design-partner narrative for the Ivertiq Full-Stack AI Harness: LC–MS / LC–MS/MS instrument PDFs, sovereign on-prem execution, and human experts on every ambiguous or conflicting case. Not a guarantee of results in every lab.
2,000+
PDFs in a batch window
Reference scale on workstation-class dual-GPU hardware for monthly-style processing loads.
On-prem
Sovereign by design
Forensic evidence, Lab IDs, and chemical results stay inside the customer environment.
Metrics are illustrative POC/MVP selling aids pending validation under each customer’s SOP, QA protocol, input formats, and hardware.
The problem
In forensic drug analysis, LC–MS and LC–MS/MS (and in some workflows GC–MS) are the analytical gold standard: liquid or gas chromatography separates a complex sample into ingredients; mass spectrometry characterizes each ingredient and compares its spectrum against reference databases.
Even so, the instrument does not issue a finished judgment. For each ingredient it typically emits a measured spectrum plus database candidates — often HIT 1 and HIT 2 from two libraries — with similarity / confidence scores. Those scores are useful starting points, not sign-off. High-confidence candidates can still be wrong or ambiguous when peak patterns, impurities, noise, or retention-time context disagree. Labs therefore require expert visual inspection of the equipment-generated PDFs before findings are recorded and reported.
As volume rises (thousands of PDFs per month in some settings; tens of thousands of samples per year), adding headcount alone does not scale. Fatigue increases QC risk. Sensitive evidence, Lab IDs, and chemical results cannot be casually uploaded to public cloud AI.
Instrument confidence is not a final answer. The goal is not to replace forensic experts — it is to accelerate disciplined review, surface conflicts early, and keep every accountable decision human-gated.
What makes the PDF hard
- Crowded evidence — each ingredient may present multiple spectrometric diagrams that must be read together with labels, retention time, and text HIT results.
- Close or conflicting HITs — two strong database matches can both be wrong, or only one may be correct once impurity peaks are excluded.
- Mandatory expert gate — similarity scores alone are not sufficient for sign-out; humans still certify the read under lab SOP / QA.
Phase 1 scope (illustrative)
01
Ingest
Instrument PDFs / exports and controlled mapping references.
02
Extract & map
Lab ID, HIT 1 / HIT 2 candidates, similarity cues, and spectrum-linked evidence → controlled drug mapping rules.
03
Gate
Consistent mapping → structured output; conflict, close HITs, or low-quality spectra → expert review.
04
Collate
Optional PPT assembly from LIMS export + selected sample photos.
Why it proves the Harness thesis
The same stack pattern that matters here — local execution, structured evidence, deterministic conflict gates, and HITL — is why generic public-cloud copilots are the wrong default for forensic PDFs.
- Sovereignty — forensic data stays local; no public LLM API path for evidence.
- Deterministic gates — automation stops when Hit1/Hit2 disagree; experts own exceptions.
- Auditability — structured outputs and review paths support defensibility.
- Expansion path — same stack pattern extends toward clinical/health-screening LC-MS review and other lab document workflows.
What we do not claim
- Replacement of forensic expert judgment
- Guaranteed accuracy or elimination of all QC risk
- Coverage of every forensic subsidiary workflow on day one