Selected work / Engineering in context

From autonomous repair
to AI that joins the room.

Agent swarms, a conversational avatar, a Slack responder, retrieval tooling and products with paying users. Different interfaces. The same attention to the system behind them.

01 / Knight CTO · Autonomous delivery

A repair needs
an independent verdict.

An autonomous acceptance swarm coordinates the work between agents that find defects, repair code and test the result. Boris built this system through Knight CTO client work.

The engineering decision

Give discovery, repair and retesting separate responsibilities and isolated workers. The fixer produces a candidate; another worker evaluates it against the original acceptance criteria. Temporal coordinates the workflow, while PostgreSQL retains state, task contracts and evidence.

What the work demonstrates

A documented bounded proof completed a real repair of a disposable notes application: source revision, rebuild, independent browser retest and promotion. The result establishes that workflow. Broader continuous operation and arbitrary-application reliability remain under development.

Agent harnessesHybrid loopsTemporalPostgreSQLKubernetesIndependent acceptance

Start with the intended behavior.

A finder investigates the authorized application against its original requirements. It records the observed defect and supporting evidence, giving the repair worker a concrete problem to solve.

02 / Knight CTO · Conversational AI

Atlas.
A voice for the knowledge.

Created Atlas, an AI avatar that connects a conversational interface with a source-backed knowledge system. The work goes well beyond putting a face on a chatbot.

Conversation is a system

Meeting integration, speech recognition, streamed answers and avatar speech have to agree on whose turn it is. The engineering includes turn state, interruption controls, grounded speech, session ownership and media teardown, with Tavus and Recall integrations.

Knowledge behind the answer

The knowledge layer ingests and reconciles sources, chunks documents, computes local embeddings and retrieves relevant passages with provenance and access checks. Conversation capture and memory make continuity an engineering concern alongside answer quality.

Evidence and scope

The project includes implemented media and knowledge components and retained indexing and retrieval exercises. Live conversation behavior is evaluated separately for turn-taking, interruptions and audible responses.

Voice agentsAvatarsRAGLocal embeddingsConversation stateSource grounding

03 / Knight CTO · Workflow automation

An agent that knows
when to respond.

Built a Slack auto-responder that connects live events, conversation context and agent judgment. Deciding whether a response is appropriate is part of the work.

The conversation comes first

The responder reads fresh thread context, distinguishes a directed request from background conversation, and gives human replies precedence. Unresolved decisions can remain with the owner while independent work continues.

Delivery has a state

Events enter a durable queue; replayed events are deduplicated. A shared send ledger and exact reply reconciliation track delivery. An uncertain send remains unresolved instead of automatically sending again. Authorship and reply controls are explicit.

Evidence and scope

The transport and workflow have been exercised locally, with real cloud events and bounded worker runs also verified. Automatic cloud reply delivery remains a separate acceptance milestone.

Slack Socket ModeAgentic judgmentDurable queuesThread contextDeduplicationHuman handoff

04 / SettleOn · Founder & CEO

From customer problem
to paying users.

Led a 10-person team spanning engineering and business development, including a tech lead and a business-development manager. Brought AI products to market and owned the work around the launch.

Product and commercial delivery

Legal AI served law-firm customers. An automated form-filling product reached several dozen paying users. The work joined customer discovery, product architecture, hands-on engineering, integration, onboarding and support.

Agents inside the business

RevNode used multi-LLM sales agents to support internal go-to-market work. It is an example of internal automation; customer adoption belongs to the legal and form-filling products.

Product developmentAI automationMulti-LLM agentsTeam leadershipCustomer delivery

05 / Open source · Evaluation

Test the evidence
behind the answer.

Built MCP Retrieval Probe, a Python CLI and MCP server for inspecting retrieval systems. It turns grounding quality into something you can investigate directly.

Questions the probes answer

Does confidence track support? Can the system abstain when the corpus cannot answer? Are relevant sources retrieved and citations linked? Calibration, abstention, embedding anisotropy, relevance and corpus coverage each expose a different failure mode.

Integration depth

Server-specific recipes target different MCP retrieval services. Additional probes inspect endpoint-count consistency and OAuth discovery. The public repository includes the implementation and probe recipes.

PythonMCPCalibrationAbstentionCoverageCitation linkage
Explore the public repository

06 / Knight CTO · Customer deployment

Useful inside
the customer’s boundaries.

Legal AI is one concrete application of the broader practice: retrieval over dense material and multi-step workflows inside a customer-controlled environment.

The deployment is part of the product

The engagement included architecture, integration, deployment and support, with customer-controlled model keys, firm-owned audit logs, zero-data-retention configuration and attorney supervision.

What carries across domains

Source grounding, explicit permissions, human supervision and operational evidence matter wherever an AI system handles consequential work. The architecture must reflect the actual customer workflow.

RAGAgentic workflowsAuditabilityCustomer integrationSupervised AI

Have a difficult system in mind?

Let’s build what
comes next.

Talk to Boris

Or boris@knightcto.com