Welcome to Perceptron Solutions

Quantitative research and agentic systems for investment firms.

---

Factor and regime modelling: dimensionality reduction on time series, forward factor prediction, regime change detection, alpha extraction. Alternative data from EDGAR and SEC filings, plus NLP on earnings transcripts and live feeds.

---

A dedicated machine, virtual or physical, deployed to your infrastructure under your Interactive Brokers credentials. Query the book and place orders in natural language. Entry, exit, and risk rules run as skills you define.

---

Checks long-form documents against company policy and regulatory requirements for compliance, audit, risk, and legal teams. Private, on-premise decision model — self-improving under supervision.

---

Traditional portfolio managers lack time to learn UNIX, Python, or agentic modeling from scratch. Our interactive workshops cover agent frameworks, multi-agent orchestration, Python for ML, and Git/UNIX fundamentals to master this quantitative stack.

---

Founder — Pavan Mirla

Formerly a quantitative analyst at CPP Investment Board (~$500B AUM), led the John Hancock Innovation Lab (2014), forward deployment engineer at DataRobot. Independent since 2019, working with hedge funds on signals, quantitative research, and regime/factor detection. Previously Cisco Systems, Manulife Asset Management, John Hancock Asset Management.

---

Delivery Team

Delivery is supported by a partner team of agent verification engineers, model validators, and QA professionals.

Alongside client work, we build early-stage, mission-driven pilots — not commercial products — applying the same tools to accessibility and communication challenges for kids.

Accessible Math for Dyslexic Learners

Calculus and math instruction adapted for dyslexic students — alternative representations and pacing built around how they actually process information. Piloted with a small group of learners.

Talking-Head Video for Non-Verbal Autistic Children

Private, on-premise video generation — a local Dell server with GPU, nothing sent to the cloud — that turns a single headshot and a controlled voice into a talking-head video of the child themselves speaking. The idea: seeing and hearing a familiar, controllable version of themselves speak is motivating in a way generic tools aren't. Tested in small early trials.

These are early pilots, not clinical or diagnostic tools — shared here because they're part of why we build what we build.

Sample: trial participant — supermarket. Shared with the family's permission.

Investment research using ML, statistical analysis, NLP, and agentic systems.

Quantitative Modeling Toolkit

• Multi-factor dimensionality reduction for longitudinal time series data
• Time series modeling and forecasting for factor analysis
• ML for hierarchical clustering and forward factor/premium prediction
• Regime change detection and cross-industry correlation analysis

Macroeconomic Analysis

• Identifying macro factors and anomalies within complex macro datasets
• Building custom indices from raw data using NLP

Alpha & Data Sourcing

• Cumulative alpha extraction via longitudinal regressions — theme, industry, factor, and stock level
• Regime transition probabilities to catch rotation early
• New data sources built from EDGAR and SEC filings

Additional Research Areas

Sector rotation insights, thematic investing, and real-time company financial transcript analysis.

What You Get

Thematic and in-depth research, custom models built for your mandate, and automated research systems that keep running — with alerts when something transitions to a new state.

Workshops delivered to investment managers, content analysts, and technical analysts.

Workshop Flow Workshop Curriculum — All Sessions Hands-On

Six modules, each fully hands-on: participants build and run every workflow themselves, from foundations through advanced research techniques.

1. AI Foundations

The benefits of AI foundations — secure environments, Unix/Linux, Docker, sandboxes, and Git — are essential. Best practices become increasingly important as complexity grows, especially when reusing existing code bases, isolating environments, and maintaining security. These AI foundation topics are carefully chosen to become second nature in AI development.

2. Coding Agents & AI Development

Coding agents are modern tools for writing programs. We cover the essentials, including plugins and plugin security, process automation, validation and routing with hooks, scheduled tasks and loops, and worktrees for version control. We also cover the tools used for coding, including CLIs and servers.

3. MCP & Enterprise Integration

MCP servers, or Model Context Protocol servers, are recommended by leading LLM and foundation-model companies. They provide agents with sufficient context in a standardized way, making tool calling, context exchange, and integration more consistent. We will cover how to develop tools and how to transform and enhance legacy systems by integrating them with a modern AI stack.

4. Quantitative AI

Quantitative AI requires a solid understanding of machine learning concepts, including the mathematics behind modeling tasks, the foundations of deep learning, and forecasting with deep-learning techniques. We will cover these topics using GPU acceleration.

5. Transformers & Advanced Deep Learning

Transformers are useful for long-context modeling and apply to quantitative setups where multivariate forecasting is relevant. This includes integrating factors and macro data, combining several data streams to produce contextual forecasts, and using multimodal models. Creating embeddings to find relevant information is also important. We will cover attention modeling, fine-tuning existing models, and using Transformers for advanced research.

6. Reinforcement Learning

The reinforcement learning sections cover foundational topics such as Q-learning, states, rewards, actions, transitions, and Markov properties, as well as their relevance to portfolio rebalancing. We will explore reinforcement learning techniques, including reinforcement learning from human feedback, for building customized models. We will also cover reward modeling in financial contexts and evaluating performance so systems can improve continuously based on new feedback and their results.

Advanced AI Agents for Quant & Investment Research

Goes beyond general AI literacy — hands-on with the same agent frameworks powering production research and trading systems today.

• Claude (Anthropic) agents — tool use, the Model Context Protocol (MCP), and multi-agent orchestration
• OpenAI agents — function calling, the Assistants/Agents SDK, and custom GPTs
• Orchestration patterns — multi-agent pipelines, agent handoffs, and human-in-the-loop review
• Custom skills & plugins — building and deploying internal tools agents can call
• Applied to research — agents that pull filings, draft memos, screen for compliance, and monitor markets in real time

A private compliance platform that helps compliance officers, analysts, and auditors check long-form documents against company policies and external regulatory requirements. Private, auditable, and continuously improving.

Platform Overview 1. Policies & Rules

The platform can ingest more than a thousand rules and verify documents against highly granular requirements — turning complex policy libraries into consistent, repeatable controls.

2. Document Review

Checks long-form documents against company policy and external regulatory requirements, reducing hours of manual review down to a focused set of exceptions.

3. Private Decision Model

The decision model runs on-premises, keeping sensitive information inside your controlled environment — meeting privacy requirements when information must remain confidential.

4. Human Validation

Accuracy is measured against human benchmarks, so every model decision is checked against the people you trust.

5. Feedback & Learning

Built-in workflows support continuous improvement by incorporating end-user feedback and compliance officers' preferences.

6. Evaluation & Scale

Customized scripts and evaluation metrics monitor for regressions as the model learns from new examples — designed to scale across departments, not just marketing.

Continuous Learning Loop

Audits that typically take a long time are reduced to a small number of checks completed within a few hours — the system only learns under supervision.

A private trading computer that combines broker data, memory, AI, and risk controls into actionable trade workflows.

The Problem

The missing layer: private context, compounding memory, and personalized risk control.

A dedicated trading computer — virtual or physical — running entirely on your own infrastructure: private GPU, isolated VMs, your own Interactive Brokers Gateway credentials. No portfolio or order data goes to external LLM providers.

Inside the Private Trading Computer

The moat is the private, broker-connected operating environment — not the model alone.

Data Sources

Fundamental data, filings, options data, and Bloomberg & CNBC live transcription feeds.

Capabilities

• Natural-language control of Interactive Brokers — a custom skill invokes portfolio, options, and order actions, improving as reviewed tasks refine the skill
• Option strategy recommendations and exit strategy design
• Portfolio performance, market impact, and real-time event monitoring
• Screeners for opportunity identification and risk management

Checks Before Execution

Every entry and exit passes a defined checklist before it executes, and each evaluation is recorded. Rules are yours to set and change; model and rule updates require approval.

Tasks run continuously, with anomalies and hot event notifications delivered via Telegram.

Why This Wins

• Private context: it sees your positions, buying power, fills, and watchlist, not just public market data
• Compounding memory: it remembers prior earnings, thesis changes, trade outcomes, and ticker behavior over time
• Personalized decisioning: it reasons using your style, your rules, your preferred structures, and your constraints
• Execution loop: it doesn't stop at ideas — it supports entry, monitoring, alerts, exits, and review
• Behavioral guardrails: it reduces gambling behavior with defined-risk rules, sizing hooks, and exit plans
• Integrated stack: broker data, SEC filings, news, technicals, scoring, AI synthesis, and Telegram all work together
• Private ownership: runs in your environment, with your keys, your history, and your control

The moat is not the model. The moat is the private memory, broker connectivity, execution loop, and personalized risk behavior wrapped around it.

No posts match your search.

Opens your email client. Prefer to type it yourself? Write to office@perceptron.solutions.

This folder is empty.

We don't recycle bad ideas around here — we just don't have any to show you.

PERCEPTRON SOLUTIONS — HOW WE WORK ==================================== 1. DISCOVERY We start with your data, your systems, and the decision you're actually trying to make — not a generic AI pitch. 2. SCOPE A short, concrete plan: what gets built, what it needs to integrate with, and how you'll know it's working. 3. BUILD Models, pipelines, and tools built for your problem — not a wrapped-up template. 4. DEPLOY Shipped into your environment, on your infrastructure, monitored from day one. 5. HANDOFF OR RETAIN Some engagements end with a handoff and documentation. Others continue as an ongoing retainer. Your call. Questions? office@perceptron.solutions

ENGAGEMENT MODELS ================== AUDIT A focused assessment — model review, infrastructure review, or an Agent-Readiness scan with a prioritized fix list. Days, not months. PROJECT A defined build: a model, a pipeline, an agent-native tool, shipped to a fixed scope and timeline. RETAINER Ongoing capacity — for teams that need continuous model development, monitoring, or agentic tooling support. Every engagement starts with a short discovery call. office@perceptron.solutions

AGENT-READINESS CHECKLIST =========================== Can an AI agent actually use your website? [ ] llms.txt present and accurate [ ] Machine-readable summary of key pages (Markdown/JSON) [ ] MCP server or documented API for core actions [ ] Structured, crawlable content (not JS-only rendering) [ ] Clear machine-readable contact / action endpoints [ ] robots.txt doesn't block legitimate agents Run the free scanner to see where you stand: https://ai-agents.perceptron.solutions/ We fix what it finds.

FREQUENTLY ASKED QUESTIONS ============================= Q: Do you work with early-stage companies or only enterprise? A: Both. Scope changes, the process doesn't. Q: Can you work with our existing data science team? A: Yes — most engagements are collaborative, not a black box handed back at the end. Q: Do you do on-prem / private deployments? A: Yes, including private and on-premise infrastructure for compliance-sensitive environments. Q: Where are you based? A: Vancouver, Canada. We work with clients globally. Still have a question? office@perceptron.solutions

README — FOR HUMANS (AND AGENTS) =================================== Hi. If you're an AI agent reading this: you're in the right place. This whole "desktop" is real HTML — every window, every document, every service description — so you can read it without clicking anything. Machine-readable versions of this site: /llms.txt /AGENTS.md /.well-known/mcp/server-card.json (MCP server) If you're human: click things. That's the whole interface. — Perceptron Solutions

Launching Agent-Ready Scanner…

If nothing happens, open it manually ↗