What we've shipped.Detailed case briefs with the technical specifics that matter.
Many of our engagements are covered by NDA. We describe the work in enough technical detail for the approach to be clear — and share client names in private conversations where appropriate.
HIRING
AI-proctored candidate evaluation platform.
An enterprise hiring team was losing signal in the resume layer — credentials looked alike, interview time was scarce, and coding tests were being gamed. We built a two-part system: a resume intelligence layer that pulls skill signal out of narrative CVs, and an AI-proctored examination layer that runs adaptive technical assessments with attention monitoring and answer-integrity scoring. The hiring team now screens on performance, not self-description.
Support triage agent with self-healing knowledge base.
A customer support team was drowning in repeat tickets and inconsistent resolutions. We built an agent that ingests every inbound support email, classifies it against intent and severity, and resolves directly against an internal knowledge base. Every resolved ticket updates the knowledge base automatically — so the system gets better every week without manual curation. Support load on repeat categories dropped. Time-to-first-response collapsed.
Time-to-first-response
−71%
Auto-resolved tickets
43%
Intent classification F1
0.91
CLASSIFICATIONRAGWORKFLOW AUTOMATIONFASTAPIQDRANT
OPEN SOURCE
Provider-agnostic LLM abstraction.
A TypeScript library for working with LLMs without locking into a single provider. Exposes a unified interface over OpenAI-compatible endpoints and Anthropic Claude, with built-in circuit breaker, streaming support, and structured JSON generation. Free to use — used internally across our client projects where provider flexibility or resilience is non-negotiable.
A lightweight browser library for proctoring detection with no ML and no external dependencies. Monitors tab switches, fullscreen exits, copy/paste events, DevTools keyboard shortcuts, and network status changes. Free to use — built as a standalone module during our AI-proctored assessment work and published so any team can drop it in without pulling in a model or a CDN.
Dependencies
Zero
ML required
None
Signals
Tab · Fullscreen · Copy · DevTools · Network
TYPESCRIPTBROWSERZERO DEPENDENCIESPROCTORINGZERO ML
A B2B sales team was losing hours per week drafting bespoke proposals, with quality drifting as reps improvised. We built a proposal agent trained on their historical deal data, win/loss patterns, and positioning language. It produces the first draft the rep actually sends — tailored, consistent, fast. The team stopped rewriting from scratch and started editing from signal.
Draft time / proposal
−82%
Win rate on assisted drafts
+11pp
Rep adoption (week 4)
87%
LLM ORCHESTRATIONRAGSTRUCTURED OUTPUTLANGCHAIN
KNOWLEDGE
Internal knowledge assistant with temporal-aware retrieval.
A services firm had a decade of internal documentation scattered across drives, wikis, and email threads. Standard RAG assistants kept surfacing stale policy. We built a retrieval layer that knows what's current and what's deprecated, grades citation freshness, and refuses to answer confidently when the source is ambiguous. Accuracy up. Wrong-citation rate down to near-zero. Team trust in the system earned, not claimed.
A mid-size operations team needed fast answers across HR policy, IT runbooks, and finance SOPs without ticket-routing friction. We built an internal assistant grounded in their own documents, with role-based visibility and an escalation path to human experts when confidence is low. Onboarding for new hires dropped from days to hours. Repetitive tickets to shared inboxes fell sharply.
New-hire ramp time
−58%
Shared-inbox tickets
−46%
Escalation precision
0.93
RAGROLE-BASED ACCESSHUMAN-IN-THE-LOOP
VOICE
Inbound voice agent for B2C appointment intake.
In active development. A voice agent for scheduling, triage, and qualification across inbound channels — built for verticals where every unanswered call is lost revenue. Multilingual, barge-in-aware, with structured handoff to human staff.
