The cancer treatment management tool is designed to assist doctors in efficiently managing patient treatments. The application integrates with existing healthcare systems to provide a seamless experience for doctors and patients. It offers features for tracking treatment progress, managing medication schedules, and accessing patient history. By providing a user-friendly interface and advanced analytics, the tool helps doctors make informed decisions and improve patient outcomes.
Codebridge
AI Systems & Custom Software Development
Codebridge
3.9★★★★0 reviews8 Portfolio
Codebridge Portfolios
Knowledge Cloud is a knowledge management platform with an AI research assistant, built for the Tax and Legal team of a Big-4 firm. Practitioners ask natural-language questions, receive answers grounded to primary authorities, compare treatment across US states, and route every AI output through expert review before it becomes a client memo. CHALLENGE The team needed AI research they could trust for regulated work. Off-the-shelf tools hallucinate on tax and legal questions. Existing knowledge platforms silo content across systems that do not reason across jurisdictions. The team needed a system that combined a structured knowledge base with an AI research layer that cites its sources, flags its confidence, and hands every output to a senior practitioner for verification. SOLUTION Codebridge built an AI intake and analysis layer on top of the client's existing knowledge base. Practitioners type tax and legal questions. The AI returns a synthesized answer grounded to primary authorities with inline citations, a confidence indicator, and a source panel. Every answer passes through a review queue where a senior practitioner approves, edits, or rejects it. Cross-jurisdictional comparison surfaces material differences across US states. The platform includes retrieval-augmented generation over 229,000 primary authorities, expert-verification workflow, and confidence scoring on every output. RESULT Knowledge Cloud cuts jurisdictional research from hours to minutes. Every answer traces to a primary authority. Practitioners approve, edit, or reject AI output before it reaches a client. The platform tracks agreement rate, override rate, and source coverage.
Lispr is a voice dictation app for macOS and Windows built by Codebridge Technology, Inc. It captures speech and inserts transcribed text at the cursor in any application, in 99 languages, with no account required to dictate and no audio stored by default. The free plan covers 60 dictation minutes a week, and Pro is $79 a year or $7.99 a month. CHALLENGE Apple's built-in dictation barely supports non-English languages, including Ukrainian, and skips others entirely. Standard dictation tools add a 1.5-2.5 second delay between speech and text, breaking the user's train of thought. Every existing option also demands an account, an email, and data retention the user never asked for. Codebridge needed an app fast enough to feel like typing, private enough to need no sign-up, and fluent enough to serve the speakers other tools ignore. SOLUTION Codebridge built the client natively in Swift on AppKit, AVFoundation, and CoreAudio, and wrote a custom Opus encoder that uploads audio in 20ms packets while the user is still speaking, rather than waiting for the recording to end. The transcript lands at the cursor the moment the user releases the key. That change cut perceived latency from 2,500ms to roughly 300ms. On the backend, Cloudflare Workers and Durable Objects handle state, and recognition routes to the nearest available region to keep latency steady. AI translator agents localized the interface into 40 languages, with each translation scored automatically and reviewed by a human before release. Codebridge applied the same agent-assisted workflow to code review and competitor research, letting a 5-person team ship an Apple-notarized production app in two months. RESULT In its first three weeks live, Lispr processed 46,011 dictations across 29 countries, led by Ukraine, Indonesia, and Germany. The macOS download is about 6 MB and about 17 MB installed, and the Windows installer is about 14 MB. Dictation needs no account, and the app runs on a single Cloudflare Worker at about 1ms of CPU per dictation. Windows has shipped: Windows 10 and up, 64-bit, code-signed through Microsoft Azure Trusted Signing with Codebridge Technology, Inc. as the verified publisher.
RecruitAI is a production-grade AI-assisted recruitment platform Codebridge built for a US technology enterprise (1,000+ employees) processing 1,500–3,000 engineering applications monthly. The Problem: senior engineers spent 200–400 hours/month reviewing test submissions manually. Keyword-based auto-screening let candidates bypass filters. Interview-to-offer ratio of 12% signaled mismatches caught too late. What We Built: A five-agent system on LangGraph and LangChain. Intent Detection Agent: proprietary 0–100 Relevance Index (career progression, technical fit, open-source contributions, publications) Screening Agent: CV validation with RAG grounded in internal hiring standards Assessment Agent: personalized tests with marker questions to detect AI-generated responses Interview Agent: synthesizes Fireflies.ai transcripts into structured candidate profiles Onboarding Agent: Just-in-Time learning paths from Confluence for new hires Agents act autonomously only when confidence exceeds 90%. Final rejections require human review. The React Recruiter Dashboard surfaces reasoning chains for transparent override. Hierarchical LLM routing cuts per-candidate cost by 40% to $1.50–$3.00. Results: Full-cycle hiring time: 24 days ? 10–12 days (50% reduction) Candidate response time: ~24 hours ? under 2 minutes Interview-to-offer ratio: 12% ? 38% Engineering test-review time: 200–400 ? 100–150 hours/month Sourcing coverage: 3–5 ? 20+ platforms 24/7 global availability Tech: Node.js/TypeScript, React, PostgreSQL, GCP, LangChain, LangGraph, LangSmith, LLM-agnostic (Claude, GPT, Gemini).
Codebridge built a multi-agent AI system that automates sales pipeline work for a US-based B2B professional services company. It runs routine outreach and early-stage qualification across LinkedIn and email at scale, so human SDRs spend their time on high-intent prospects. CHALLENGE The client managed over 100 LinkedIn and email accounts by hand. Responses lagged across fragmented channels, and lead history sat scattered across platforms and CRM notes. The team could not sustain personalization at scale, and existing automation tools produced formulaic messages that put sender reputation at risk. The system had to respect anti-spam behavior and B2B communication standards. SOLUTION Codebridge mapped the workflow and set conservative confidence thresholds for every automated decision. A modular service-based backend with a central orchestrator syncs 100+ accounts every 5 to 15 minutes. A RAG layer grounds responses in verified company-specific knowledge. A three-stage pipeline of Context Analyzer, AI Humanizer, and Pattern Breaker removes bot-like message patterns. Intent detection classifies leads and requires 90% confidence to disqualify a lead. An autonomous CRM assistant runs follow-ups and early-stage nurturing. Gemini runs fast analysis and short-form generation, Claude Opus 4.5 takes deep reasoning and long-form content, and the Perplexity API pulls live research. Stack: Python 3.11+, FastAPI, PostgreSQL, Docker, HeyReach, Kommo (amoCRM), Calendly, Teams. Four people shipped it in one month for $20,000. RESULT Response time fell from about 24 hours to under 2 minutes. Qualified meetings rose 30% through intent-based prioritization. Time to first meeting fell from 1-2 weeks to 2-3 days. Pipeline velocity improved 30% at early sales stages. The system generates 500K+ personalized messages and frees 20K+ hours of sales time per month.
Tutorai is a web-based real-time AI tutoring platform with 3D avatars, built for a European edtech startup. Students hold live voice conversations with animated tutors across English, Science, and Life Coaching tracks on a shared interactive whiteboard. CHALLENGE Existing AI tutoring ran on a record-and-play audio loop with 3 to 5 second delays between student input and response, which broke conversational flow. The early prototype used D-ID's SaaS avatar service at $32.33 per tutoring hour, so the business model did not hold at scale. Static content formats gave students no feedback loop. SOLUTION Codebridge replaced the SaaS avatar dependency with custom WebGL 3D avatars featuring native lip-sync and multiple tutor personas. The voice pipeline runs Whisper speech-to-text, an LLM with session context, TTS, then avatar lip-sync animation. A hybrid AI strategy splits the workload between GPT-5 mini for lesson generation and OpenAI Realtime-mini for voice interaction, so the team tunes cost and responsiveness independently. A RAG layer anchors responses to curriculum materials. The tutor references PDFs, images, and homework photos that students upload mid-session. Transcripts and whiteboard state persist for session recovery. Stack: Azure OpenAI, OpenAI Whisper and TTS, Node.js, FastAPI, React, WebGL, Azure Kubernetes Service, Azure SQL, Azure Managed Redis, Azure Key Vault, Daily.co, Stripe. Five engineers cover project management, backend, AI/LLM, 3D art, and DevOps. RESULT Cost per tutoring hour fell from $32.33 to $1.15, a 96% reduction. Annual avatar cost fell from about $24,984 to about $1,049. Speech start latency dropped to under 1 second and average response time to under 2 seconds. Whiteboard sync completes in under 500ms. Session recovery is automated for 100% of lessons. The MVP runs 24/7 globally with no dependency on human tutor scheduling and is expanding into more subjects and mobile.
RadFlow AI is a HIPAA-compliant radiology workflow assistant for a Tier-1 imaging network of 12 centers across three US states. It integrates with existing PACS, ranks worklists by malignancy probability, renders studies in the browser, and keeps a radiologist in the loop on every finding. CHALLENGE Radiologists split time across a PACS viewer, a separate AI interface, and dictation reporting, and one-third of reading time went to non-interpretive tasks. Chest CT studies loaded in 8 to 12 seconds at rural satellite sites. The prior commercial AI returned 4.1 false positives per scan, so radiologists dismissed its findings without review. Scan volume grew 22% a year against flat headcount; turnaround ran 15% past contractual SLAs; and the system had to meet HIPAA/HITECH, align to IEC 62304, and stay compatible with an FDA 510(k) pathway SOLUTION Codebridge built a unified diagnostic workspace with five modules: AI triage worklist, WebGL 2.0 viewer with toggleable overlays, clinical oversight dashboard, immutable audit and explainability logs, and an integration layer. Progressive DICOM streaming over DICOMweb and adaptive bandwidth compression hold initial render under 400ms. A 3D Feature Pyramid Network on a ResNet-50 backbone reads each study in about 47 seconds, with a false-positive reduction network and a longitudinal prior-study comparison module. Grad-CAM saliency maps explain each finding, and one-click adjudication feeds a shadow-mode retraining loop. The stack is React, OHIF, PyTorch, MONAI, NVIDIA Triton, FastAPI, and AWS EKS, delivered under IEC 62304-aligned CI/CD, HIPAA Safe Harbor de-identification, TLS 1.3, and a design history file for a future 510(k). Eight engineers shipped the platform in 24 weeks. RESULT CT reading time fell from 15.2 to 9.4 minutes and P95 turnaround from 6.2 to 3.8 hours. Sensitivity on sub-4mm nodules reached 96% against a 93% threshold, and false positives per scan fell from 4.1 to 0.4. Satellite load times fell to 0.4-0.9 seconds, and radiologist trust rose from 27% to 89%. Uptime is 99.97% over nine months, and a double-blind study on 2,400 scans confirmed the claims. Estimated annual operational impact is about $2.1M.
Codebridge built a computer-vision system that measures stock on racks and floors across a 127-depot US distribution network and reconciles it against the warehouse management system in near real-time. CHALLENGE More than 100 warehouses had no stock visibility between cycle counts. Empty pick-faces went undetected until a picker arrived, and mis-stows surfaced days later through pick failures. Manually keyed inbound dimensions produced cubing errors in slotting and load planning. Raw video could not leave the site, depot links were low-bandwidth, and capture could not slow the floor. The core difficulty was metrology: turning pixels into defensible centimetre measurements with honest error bars. SOLUTION Codebridge built a five-layer edge pipeline. Instance segmentation localises items, monocular depth recovery uses calibrated ground planes and reference objects, dimensions come from segmented silhouettes, every measurement carries a confidence envelope, and the output reconciles against WMS state. Inference runs on-site on NVIDIA Jetson or x86 GPU nodes, so only measurements and variances leave the depot. Python containers, PyTorch, ONNX Runtime and TensorRT, RTSP ingest, store-and-forward sync, REST cube write-back, React/TypeScript console, Azure with regional data residency. Low-confidence results go to human review. RESULT Sub-centimetre accuracy on warehouse cartons, validated against caliper ground truth. Manual keying of cube data dropped substantially. Mis-stow detection moved from days to the same shift. The system flags an empty pick-face before a picker walks to it, cutting stockouts at the face. Better cube data reduced rework in slotting and load planning. Each depot runs as an independent edge unit, so adding sites means adding hardware, not re-platforming.